88Z AI / FLASH
Model launches, product updates and notable industry moves, grouped by date. Each item opens a full article on this site.
Google announces that Gemini app users in the U.S. can now connect their Google Wallet accounts to the app. This integration enables Gemini to provide more targeted responses and financial advice based on users' actual spending habits, marking a deeper expansion of personal AI assistants into financial management.
Source: blog.google
Read on 88z.aiAs frontier-scale models grow, their weights no longer fit into the GPU memory of a single machine. NVIDIA demonstrates how to use the vLLM framework to serve the GLM-5.3 model across two NVIDIA DGX Stations, overcoming hardware limitations for local deployment of ultra-large models.
Source: developer.nvidia.com
Read on 88z.aiNVIDIA introduces SOMA Retargeter, leveraging human motion data as a teaching foundation by capturing coordination and timing information to address humanoid robot movement control challenges. The tool aims to simplify the mapping process from human demonstration to robot execution, enhancing development efficiency.
Source: developer.nvidia.com
Read on 88z.aiAllenAI has released XYBench on Hugging Face, designed to evaluate large models' ability to solve the "problem behind the question." This benchmark goes beyond surface-level answer accuracy, deeply examining logical deduction and root cause analysis, offering a new dimension for measuring true AI intelligence.
Source: huggingface.co
Read on 88z.aiGoogle has released Gemini 4 Hydrogen, focusing on high efficiency and low hallucination rates. The model scores 51 on the intelligence index in High mode, surpassing Argon, and reaches 54 in xhigh mode, approaching Claude Sonnet 5.5 and GPT-6.1 Sol. API pricing is set at $1.5 per million input tokens and $7.5 per million output tokens. It offers 2.5 times the speed of Gemini 3.8 Flash with an 8% hallucination rate.
Source: APPSO
Read on 88z.aiiFlytek has launched SparkAudio 1.0, its first voice foundation model trained on fully domestic computing power. The model supports multi-language and dialect recognition, offers deep audio understanding capabilities, and achieves significant performance improvements and cost reductions through three technical innovations.
Source: 讯飞开放平台
Read on 88z.aiFei-Fei Li’s World Labs has publicly opened access to its next-generation world model, Atlas. Users can now generate 3D Gaussian splats from images via the Viewfinder web app or create explicit 3D geometry using agentic workflows in Chisel. The company also released all underlying API endpoints, enabling developers to build custom tools or allow agents to interact directly with the models, with outputs including keyframes, video clips, and 3D Gaussian splats.
Source: @theworldlabs
Read on 88z.aiPerplexity has open-sourced its latest decision models, Decider v1.3 (~27B parameters) and the distilled Decider Flash v1.3 (~4B parameters). The update introduces a "locate" feature that extracts text spans with character offsets based on instructions, significantly improving accuracy in entity extraction and classification tasks. Supporting text and image inputs with a 256K context window under the Apache 2.0 license, the models are available for free on OpenRouter until November 19, 2026.
Source: @perplexitydevs
Read on 88z.aiFor tasks like scheduling with multiple valid answers, traditional supervised fine-tuning is inadequate. This briefing explains how to use Amazon SageMaker AI's serverless model customization workflow and Training Sessions APIs combined with RLVR (Reinforcement Learning with Verifiable Rewards). By building verifiers with hard and soft constraints and using the GRPO algorithm, models are guided to generate compliant, high-quality solutions rather than mimicking a single ground truth.
Source: @akshay_pachaar
Read on 88z.aixAI announces that Grok Bot on Mac now supports seven system apps: Messages, Mail, Notes, Reminders, Calendar, Location, and Contacts. Users can complete cross-app tasks via natural language prompts, such as finding an address and scheduling events. The feature is included in X Premium and SuperGrok subscriptions, requiring manual permission grants and user confirmation for sensitive actions.
Source: @bot
Read on 88z.aiAnthropic has announced a private beta for self-hosted Claude Managed Agents, allowing enterprises to deploy the agent harness, sandboxes, and session history in their own cloud or on-premises environments. This solution privatizes the orchestration layer rather than running the model locally, with model requests still directed to Anthropic or customer-chosen providers, addressing security needs for accessing internal code and systems.
Source: @ClaudeDevs
Read on 88z.aiLooki has released its self-developed multimodal model series, Loom, focusing on first-person life understanding and long-term memory. The series includes Loom Sense, Loom Moment, and Loom Memory, designed to optimize continuous perception, event segmentation, and cross-time memory retrieval. It significantly reduces inference costs and improves processing efficiency, turning fragmented observations into traceable personal context.
Source: LookiAI
Read on 88z.aiAlibaba's AI music platform HappyShrimp has officially launched on iOS and Android. Users can generate complete songs by inputting emotions or stories, support chat-to-song and video scoring, and use the new 'Make Same Style' feature to replicate creative ideas.
Source: 阿里云
Read on 88z.aiTencent Hunyuan has released Hy Realtime, a real-time voice model that enhances spoken language understanding, instruction following, and emotional expression. It supports tool calls like web search and integrates with Tencent Music for song requests. The model is live in Yuanbao voice calls and Tencent Customer Service, with API access coming soon via Tencent Cloud.
Source: 腾讯混元
Read on 88z.aiY Combinator-backed Taan released Taan v1-preview on November 12, 2026. The model accepts multimodal inputs like humming, MIDI, images, or scripts to generate music with precise timing control. The team reports a 96% instruction success rate, aiming to solve the lack of controllability in current AI music models.
Source: @mp3addi
Read on 88z.aiThinking Machines Lab has officially open-sourced the weights for its Inkling 2 Small model on Hugging Face under an Apache 2.0 license. Led by former OpenAI CTO Mira Murati, the 594-billion-parameter model supports ultra-long context and multimodal processing. This release marks the transition from closed testing to full openness, allowing developers to download and integrate the model immediately.
Source: @thinkymachines
Read on 88z.aiLondon AI lab Continuous Intelligence has launched ci-proto, its first time-series foundation model. The model ranks first on the fev-bench benchmark, outperforming counterparts from Google, Amazon, and Mistral, and is available with open weights for commercial use.
Source: @benjaminwalker
Read on 88z.aiAnthropic's official developer account released a video detailing the API launch process for new models and their accompanying features. The core advice highlights that when upgrading to new models like Claude Fable 5.2, developers should not rely solely on official benchmarks. Instead, they must rethink their agent harnesses and measure effectiveness via "cost per task" rather than "cost per token," while comparing correction needs and completion times in real-world business scenarios.
Source: @ClaudeDevs
Read on 88z.aiApplied Materials CEO Gary Dickerson stated that approximately 60% of capital expenditure from AI hyperscale cloud providers is directed toward the semiconductor sector. Over 85% of this spending focuses on wafer foundries, DRAM memory, and advanced packaging, which are projected to be the fastest-growing segments in wafer fab equipment by 2027.
Source: 36kr.com
Read on 88z.aiFlinn AI filed lawsuits in Delaware against Anthropic and OpenAI, alleging infringement of five patents related to attention-based neural networks.
Source: 36kr.com
Read on 88z.aiGoogle Labs has released five new Google Flow tools, designed by filmmakers and visual artists to help users bring their storytelling ideas to life.
Source: blog.google
Read on 88z.aiGoogle has launched Gemini 4 Hydrogen to replace the Flash series, with benchmarks indicating it outperforms the more expensive Argon model. While its cost-performance ratio approaches that of Claude Sonnet 5.5, its per-task cost remains 2.5 times higher than GPT-6.1 Sol, and it still trails in agentic coding capabilities.
Source: the-decoder.com
Read on 88z.aiAnalytics platform Amplitude has launched Wave, a proactive product agent built primarily on OpenAI models. The tool identifies what to build, assists with coding, and measures impact, enabling teams to ship product improvements in just 1–2 days.
Source: openai.com
Read on 88z.aiAmid the global decline of giant kelp forests due to ocean warming, scientists are employing Google AI technologies for ecological restoration. These underwater forests are vital for coastline protection and marine habitats, marking a significant tech-driven approach to environmental conservation.
Source: blog.google
Read on 88z.aiGoogle has announced a $10 million Energy Impact Program in West Virginia to fund home repairs and efficiency upgrades. The initiative is open to local residents, and eligible users can apply now.
Source: blog.google
Read on 88z.aiGoogle has integrated the Gemini assistant into the Chrome browser, allowing users to get quick answers, article summaries, and explanations of complex topics directly via the sidebar. This feature aims to enhance browsing efficiency by enabling information filtering and comparison without leaving the current page.
Source: blog.google
Read on 88z.aiSYNAPUFF, an interaction intelligence company for AI-native terminals, completed a seed round of several million yuan, invested by FIRST RULE VENTURES and Pudong Venture Capital. The team proposes the "P Framework" to enhance agents' contextual understanding and dynamic behavior adjustment, addressing limitations of traditional command-response models through non-verbal probing and feedback loops for more natural interactions.
Source: 36kr.com
Read on 88z.aiChrome's developer team has released Modern Web Guidance 1.0, a toolkit that injects 200+ expert-written guides and Baseline browser-support data into AI coding agents like Cursor, Copilot, and Antigravity, steering them to write modern, standards-aligned HTML, CSS, and JavaScript by default. The team's own evals found that even the latest frontier models fail more than 40% of edge cases on modern web best practices.
Source: @ChromiumDev（X）
Read on 88z.aiBaidu has released its first post-trained model built on ERNIE 5.5 Flash Base: 280B total parameters, 14B active, a 1M-token context window, with weights to be open-sourced by November 25.
Source: @Baidu_ERNIE
Read on 88z.aiThe latest DesignArena ranking places MiniMax H3-Image second among open-weight image models, averaging 15 seconds per generation, making it the fastest model in the top 3.
Source: @DesignArena
Read on 88z.aiAnthropic has announced its initial public offering, having filed a registration statement with the SEC, with Goldman Sachs and Morgan Stanley involved in distributing the preliminary prospectus.
Source: @AnthropicAI
Read on 88z.aiAnthropic has launched an office productivity suite for Claude, featuring spreadsheets, whiteboards, and data analysis with real-time collaboration and third-party integrations. Microsoft's stock fell following the announcement, amid concerns over potential disruption to its long-standing Office business.
Source: APPSO
Read on 88z.aiByteDance has launched Seedance 2.5 flash, a lightweight video generation model that doubles generation speed while maintaining quality close to the flagship version and significantly reducing costs. It supports multimodal inputs and longer video generation, and is now rolling out on platforms such as Jimeng and LibTV.
Source: 火山引擎
Read on 88z.aiAlibaba has launched the preview of Qwen4-Max, debuting on the Qianwen AI Platform Token Plan. The model features comprehensive upgrades in world knowledge, perception, coding, and agentic capabilities, excelling in complex programming, professional office tasks, and design. Internal tests highlight strong performance in 3D modeling and human-computer interaction.
Source: 千问AI平台
Read on 88z.aiSeedance 2.5 Flash has been globally launched on LibTV, accessible to all users. The model focuses on faster generation speeds and lower costs, supporting video creation from 4 to 30 seconds in 480p and 720p resolutions. Features include text-to-video, start-and-end frame generation, multi-material reference, audio-enabled video production, as well as video editing and extension. Pricing aligns with the Seedance 2.0 Standard edition, with 720p rates dropping to 0.10 yuan per second during the Do
Source: 哩布哩布AI
Read on 88z.aiOn 10 November 2026 at KubeCon, NVIDIA said it is contributing OpenShell, a runtime for securing and governing autonomous agents, to the Cloud Native Computing Foundation as a CNCF Sandbox project. It is also contributing to Agent Substrate, a Google-led project already accepted into the sandbox, and announced new DSX AI Factory Platform components for deploying and scaling GPU inference on Kubernetes.
Source: NVIDIA
Read on 88z.aiOn 10 November 2026, Hugging Face explained that organization admins can subscribe through Google Cloud Marketplace and attach Hub billing to an existing Google Cloud account. Setup has three steps: subscribe on the Marketplace, choose an organization you administer on Hugging Face, and wait for approval. Until the switch is complete, charges may still go to the previous payment method.
Source: Hugging Face
Read on 88z.ai88z.ai offers an OpenAI-compatible subset API that aggregates models from major providers and bills in CNY points. This guide follows the live /api.md page: how to register, create a site key, point the Base URL at https://www.88z.ai/v1, which endpoints are implemented, what the limits are, and how charges are calculated. It is not the official Anthropic, OpenAI, or Google API, and it is not the same site as 88api.ai.
Source: 88z.ai API 说明
Read on 88z.aiOn 11 November 2026, Google DeepMind introduced the Gemini Decisions API, powered by a new model, Gemini Decision 1 (gemini-decision-1). It is built for bounded judgments such as yes/no, multiple choice, and rubric scoring. One request can ask up to 1,000 independent questions about a shared multimodal input and return probabilities and confidence scores. The announced price is $0.10 per million input tokens, with output tokens billed at zero.
Source: Google DeepMind
Read on 88z.aiOn 11 November 2026, Google said the Gemini app is coming to Apple CarPlay, so drivers can use Gemini Live for hands-free conversations in the car. Officially, it is rolling out globally to all Gemini users on iOS, across every supported Gemini Live language. To start, you need the latest Gemini app on iPhone, microphone permission enabled, and at least one prior Gemini Live session on the phone.
Source: Google
Read on 88z.aixAI official documentation, Get started, walks through the full path: install the desktop app from x.ai/bot, sign in with a supported plan, hand the Bot a real first task, and schedule it as a routine once the result looks right. This article follows those official steps and notes how approvals, the Vault, and mobile fit in.
Source: xAI Docs, Get started
Read on 88z.aiGoogle DeepMind announced AlphaGenome 2 on 11 November 2026. The model conditions predictions on a specific cell type or state rather than tissue averages. DeepMind says it identifies nearly twice as many variants that change gene expression as the previous model, and it is available to non-commercial researchers through an API.
Source: Google DeepMind
Read on 88z.aiThe official Grok Bot account announced Grok Bot 1.0, positioned as AI teammates that work alongside an entire team. The product page says it is out of beta, with support for sharing bots, collaborating in Slack and Microsoft Teams, inviting bots to meetings, and bots that learn and become more helpful. Musk quoted the announcement the same day, suggesting hiring Grok @Bot as a colleague.
Source: @bot
Read on 88z.aiDeepMind has published a blog post introducing AlphaProof Code, described as a breakthrough in verifying complex software at depth; the same page also lists CodeMender, an AI agent for code security.
Source: deepmind.google
Read on 88z.aiGoogle DeepMind's official blog has published an article titled Gemini Robotics 2.5 accelerates real-world general dexterity, alongside a listing for Gemini 4 Argon for Robotics, without disclosing technical details or test results.
Source: deepmind.google
Read on 88z.ai@Re7_AI published an updated Grok Bot usage guide on 29 October 2026, saying the earlier tutorial was nearly half outdated. The new version is organised into three stages: day-one acceptance, running real tasks with ready-made templates in week one, and later quota management plus memory cleanup. The post is community advice; official first-run steps live in xAI documentation.
Source: @Re7_AI
Read on 88z.aiReplying on X to a question about Grok 4.8's timing, Musk said pre-training is finished, the RL and inference stack was rewritten from scratch in C/C++, and the model is now in RL with evaluations just beginning.
Source: @elonmusk
Read on 88z.aiOn 10 November 2026, OpenAI announced a beta for Businesses in ChatGPT. Eligible businesses can claim a profile, maintain product pricing and availability, and see aggregated mention and website-click insights, which OpenAI says do not expose user conversations. Business Agents initially serve eligible US ecommerce and online-service businesses, and business verification does not by itself guarantee permission to publish an agent.
Source: OpenAI
Read on 88z.aiOn 10 November 2026, Google said Gemini 4 Carbon is opening more widely to paying developers, enterprises, scientists, and consumers. Officially, Carbon is for longer-horizon, more demanding tasks. It shares Argon’s introductory unit price, but Google also says it uses more tokens than Argon on difficult work.
Source: Google
Read on 88z.aiOn 5 November 2026, OpenAI released GPT-6 Cyber as a security research model, with professional access through Daybreak. Capability and scope-adherence results in the announcement are vendor evaluations. Mozilla’s Firefox 159 security advisory, announced on 27 October, provides an independently readable record of some fixes, including reporters attributed to OpenAI.
Source: OpenAI
Read on 88z.aiIn a technical write-up dated 9 November 2026, OpenAI described Control Room. Research jobs carry identifiers, and inference engines check permission before producing output. After a job is stopped, related requests are denied. OpenAI says output stops within 35 seconds of an intervention decision; that figure is the company’s own, and this site has not independently measured it.
Source: OpenAI
Read on 88z.aiAnthropic updated Claude Fable and Mythos to 5.2 on 9 November 2026. Officially, they share an underlying model but differ in safeguards and verification. Announced standard prices are $8 per million input tokens and $40 per million output tokens, with a faster and more expensive Fast mode. For most everyday work, Anthropic still suggests starting with Opus or another generally available model.
Source: Anthropic
Read on 88z.aiOn 9 November 2026, OpenAI said ChatGPT Enterprise can be bought on AWS Marketplace. The listing describes one enterprise offering that brings together Chat, ChatGPT Work, Codex, and dots, billed through an existing AWS relationship, with eligible spend counting toward an EDP or PPA commitment. Pricing is by private offer, and a Marketplace purchase does not guarantee that all customer data stays in AWS.
Source: OpenAI
Read on 88z.aiGoogle's Gemini 4 Carbon model API is now live, available to Ultra subscribers. The model ranks first in multiple benchmarks, offers significantly lower task costs than competitors, and features automatic 3D model reconstruction and multi-agent collaboration scheduling.
Source: 量子位
Read on 88z.aiMiniMax has officially open-sourced its image generation model, MiniMax H3-Image, which supports text-to-image, image reference, and editing capabilities. The model can generate 4096x4096 high-resolution images within 10 seconds, operating at approximately five times the speed of mainstream market standards. It boasts the strongest instruction-following capability among open-source models and supports up to 16 reference images.
Source: MiniMax 稀宇科技
Read on 88z.aiMeta has released Muse Ember, its smallest model to date, focusing on complex reasoning and agent tasks with support for one million token context and native multimodality. The cost per single task is approximately $0.14, about one-third of Claude Haiku 5.5. While it outperforms the latter in several agent benchmarks, it still lags in real-world engineering coding.
Source: 智东西
Read on 88z.aiBaidu has released the ERNIE-5.5-Flash-Base foundation model, making its pre-training endpoint weights available. This sparse MoE model features 280B total parameters with 14B activated per token, pre-trained from scratch on 33.5T tokens. It supports a 128K training context and up to 1M for fine-tuning and inference. The weights are unprocessed by mid or post-training, making them suitable for domain adaptation, instruction tuning, and reinforcement learning.
Source: 文心大模型
Read on 88z.aiMeituan's LongCat team has open-sourced the LongCat-2.5 model under the MIT license, allowing commercial use. With 1.6 trillion total parameters and approximately 48 billion activated, it supports million-token contexts and long-horizon agent tasks. Optimized for domestic computing power, it significantly improves single-card training throughput.
Source: 龙猫LongCat
Read on 88z.aiAnthropic has released Claude Fable 5.2 and Claude Mythos 5.2, demonstrating strong autonomous collaboration and specialized problem-solving capabilities. A thousand agents built a bootable operating system in eight days, completed the formal proof of Fermat's Last Theorem in thirty hours, designed a four-layer PCB and manufactured three real units in one day, and cracked viral capsid folding in seven minutes on a single card. Benchmark performance is excellent, with Fable 5.2 offering equivalen
Source: 量子位
Read on 88z.aiAlibaba's Qoder has launched the anonymous model Qoduck, represented by a black duck icon without disclosed parameters. Praised for its speed and high success rate in long-horizon tasks and file editing, it supports complete deliverables like videos. It is currently free for individual users for a limited time.
Source: Qoder
Read on 88z.aiByteDance's Seed team has released Seedream 5.5, a new image creation model now available on Jimeng and Doubao. It supports up to 100 rounds of continuous editing with pixel-level preservation of unchanged areas, generates native RGBA transparent assets, and reduces text distortion to 0.4%, enhancing realism and design workflow efficiency.
Source: 字节跳动Seed
Read on 88z.aiThe Tencent Yuanqi platform officially ceased operations on November 9, shutting down all functions. Users must request data backups by January 9, 2027, after which data will be deleted; unused resources are eligible for refunds.
Source: IT之家
Read on 88z.aiGoogle has released the Gemini Omni 1.2 Flash video generation model with full API access. It supports native 15-second video generation, with prices reduced by 30% compared to version 1.1, costing only $0.07 per second for 720p resolution. Upgrades include multi-turn conversational editing, strong first-and-last frame consistency, reference image and audio guidance, more natural lip-syncing, and cinematic camera control.
Source: IT之家
Read on 88z.aiOpenAI has officially launched GPT-6 Cyber, a cybersecurity-specific model capable of autonomously discovering and exploiting vulnerabilities. The model achieved a 48.5% success rate in ExploitGym tests, an approximately 51% improvement over the previous generation. Furthermore, it reproduced 1,724 vulnerabilities in the Linux kernel, generating 948 usable exploits, and identified 81 unknown high-risk vulnerabilities in Firefox, some of which have been patched by Mozilla.
Source: APPSO
Read on 88z.aiBaidu has officially open-sourced the ERNIE-5.5-Flash-Base pre-trained base model, with weights now available on Hugging Face. The model offers FP8 and BF16 precision versions, is compatible with SGLang and vLLM deployment frameworks, and supports fine-tuning via Megatron-Bridge. As a sparse MoE architecture model, it features 280 billion total parameters with 14 billion activated per token, pre-trained from scratch on 33.5 trillion tokens, and supports a context window of up to 1 million tokens
Source: 文心大模型
Read on 88z.aiAnt Bailin has released Ling-3.2-flash, a native multimodal model with 560B total parameters and 25B activated per step, supporting a 1M context window. It features visual feedback self-check and adaptive anomaly recovery, allowing it to correct biases using rendering or interface results and retry after errors or tool failures. Using a hybrid attention mechanism of 7-layer KDA and 1-layer Gated MLA, it achieves an average decoding speed of 280 tokens/s for single users with long contexts under
Source: 百灵大模型
Read on 88z.aiKling 4.0 Flash is now fully available, supporting access via the official website, API, and various development tools, while also compatible with Adobe software plugins to seamlessly integrate video generation into post-production workflows.
Source: 可灵AI
Read on 88z.aiSand.ai has completed a $200 million A+ funding round, co-led by C Capital and Wangsu Technology, with participation from Shanghai State-owned Assets, Times Capital, and additional investment from existing shareholder Matrix Partners China. The company has raised a cumulative $300 million across three rounds in ten months, becoming a new unicorn in video generation. Founded by Cao Yue, former principal researcher at MSRA and key author of Swin Transformer, Sand.ai has open-sourced MAGI-1 and the
Source: 晚点LatePost
Read on 88z.aiOpenAI has released GPT Images 2.6, integrated into ChatGPT, Codex, and its API, with image generation speeds improved by approximately 50%. The API offers Flare and Sunburst modes, focusing on fast low-resolution output and medium-to-high resolution consistency respectively, while adding GIF animation generation capabilities.
Source: APPSO
Read on 88z.aiMoonshot AI has officially open-sourced the weights for Kimi K3.1 Pro on Hugging Face. Featuring 2.8 trillion parameters and a mixture-of-experts architecture, the model supports context windows of up to one million tokens. It ranks first among open-weight models in agent benchmarks and leads in Bash reliability compared to similar open-source counterparts.
Source: IT之家
Read on 88z.aiAlibaba Cloud has launched Qovest, an AI-powered investment research workbench designed for professional institutional investors. The platform accepts natural language prompts to automatically retrieve data, perform analysis, and generate reports. It also features a 'Research Space' for archiving insights and tracking assets at set frequencies. Built on a proprietary financial Agent framework and authoritative databases, Qovest outperforms mainstream products in quality, speed, and cost accordin
Source: 阿里云
Read on 88z.aiAnthropic has open-sourced FlashProve, leveraging zero-knowledge proofs to allow service providers to verify that responses originate from a specified model without exposing weights. Tested on Llama 4 Maverick, the system enables trusted verification despite higher costs.
Source: 机器之心
Read on 88z.aiAlibaba has launched Looogen, an AI platform for film and TV creation targeting short dramas and video production. It automates the entire workflow from script breakdown to final rendering while addressing character consistency issues.
Source: Looogen
Read on 88z.aiTencent Hunyuan has officially released Hy4, introducing visual and computer operation capabilities for the first time while enhancing instruction following and long-horizon task processing. The model ranks first among open-source models in financial and scientific evaluations and is now available on Tencent Cloud, Yuanbao, and other platforms.
Source: 腾讯混元
Read on 88z.aiSenseTime has released the official version of SenseNova 6.8 Flash, focusing on complex task execution with enhanced long-term memory, proactive interaction, and cross-task collaboration. In office scenarios, it analyzes historical customer feedback to prompt follow-ups and preserve experience. For design, it supports dynamic presentations, posters, Blender 3D, and web development, with integration into Figma.
Source: 商汤科技SenseTime
Read on 88z.aiMOSS Intelligence has released MOSS-Live, a full-duplex real-time interaction system using a front-back-end collaborative architecture. It supports simultaneous listening and speaking, interruptible dialogue, and multimodal capabilities. The system demonstrates low interruption rates and high anti-interference performance in benchmarks, enabling natural real-time voice communication alongside complex reasoning.
Source: 模思智能MOSS
Read on 88z.aiTencent Maps has launched LBSwork, a desktop AI map workspace offering intelligent support in seven areas including retail site selection. The Car Map Agent enables multi-agent collaboration and supports native ride-hailing on AI glasses and HarmonyOS cross-device continuity. The service is integrated with Yuanbao, WeChat Xiaowei, and OPPO Xiao Bu Assistant, and is open to developers via MCP interfaces.
Source: 腾讯地图
Read on 88z.aiTencent Hunyuan has released the Hyra 2.0 research agent framework, shifting focus from solving problems to building reusable tools. Featuring a three-tier architecture, it supports long-term R&D projects and has already delivered results in fields such as protein design and mathematical proofs, aiming to create compounding intelligence through tool reuse.
Source: 腾讯混元
Read on 88z.aiSenseTime has open-sourced the SenseNova Loop-9B-Preview multimodal model. Utilizing a Looped Transformer architecture, it extends effective computational depth via parameter-shared recurrent modules and incorporates an mHC mechanism to enhance feature interaction and training stability. The model matches or surpasses 27B-parameter models in code generation, office tasks, and agent workflows. Ablation studies show a 4.42-point gain on GPQA Diamond and a 13.34-point increase on TeamBench.
Source: 商汤科技SenseTime
Read on 88z.aiMeituan LongCat has released LongCat-2.5, a native multimodal model with 1.6 trillion total parameters and approximately 48 billion activated per token. Designed for long-horizon agent tasks, it enhances visual understanding and agent programming through long-sequence training and large-scale Agentic RL to improve complex code development and project building.
Source: 龙猫LongCat
Read on 88z.aiAgibot has released the GO-3 Preview embodied intelligence model, achieving scalable learning across pre-training, post-training, and deployment. The model secured first place in RoboDojo, RoboTwin, and RoboColiseum benchmarks. It uses WRAM to absorb human first-person data during pre-training, employs Genie Evolver for mixed training with real robots, simulation, and world models, and leverages Genie Master to accumulate experience in real-world settings like convenience stores.
Source: 机器之心
Read on 88z.aiOpenAI announced that starting in December, ChatGPT and Codex quotas for Plus, Pro, and Business plans will be merged and measured by tokens, with only Instant quick chat excluded from the shared pool. This move addresses recent user dissatisfaction after quota increases for some Pro and Business plans did not include Plus users. Meanwhile, Codex has tightened its multi-account policy, potentially restricting logins for devices linked to too many paid accounts. Specific quotas and exact effectiv
Source: APPSO
Read on 88z.aiKuaishou's Kling AI has completed a capital increase of approximately 19.8 billion yuan, setting a record for the largest single financing in the AI video sector. Investors include Alibaba, Tencent, Baidu, and several state-owned entities. Post-closing, controlling shareholder Lucky Labs holds 66.68%, while CEO Gai Kun retains 23.72% voting rights through special share arrangements.
Source: 智东西
Read on 88z.aiLibTV has intensified its Double 11 promotion by adding Seedance 2 Fast and Mini models to the 30-day zero-points unlimited use list, available to both new and existing members. The price for Seedance 2.5 720P has hit a record low, annual membership discounts have increased, and high-value prize draws have been added.
Source: LibTV
Read on 88z.aiPixVerse by AIsphere and OpenAI have announced a deep partnership, featuring native plugin integration with ChatGPT and Codex. Users can create videos from script to final cut using natural language within Codex, accompanied by a new creator education initiative.
Source: 爱诗科技 AIsphere
Read on 88z.aiNVIDIA has released the open-source speech transcription model Nemotron 3.5 Transcribe. Built on a MoE architecture, it supports multilingual transcription, translation, and speaker diarization, ranking among the top open-source models in various benchmarks.
Source: IT之家
Read on 88z.ai01.AI has reached a strategic cooperation agreement with Hangzhou City and Linping District, receiving hundreds of millions in combined investment from state-owned assets at both municipal and district levels. With its headquarters established in Linping, the company will jointly build an industrial AI transformation base to drive the deep application of artificial intelligence in core business scenarios such as operational decision-making, production, and supply chain collaboration.
Source: 零一万物 01AI
Read on 88z.aiVolcano Engine prioritizes commercialization, with its MaaS business generating 4.8 billion yuan in revenue and a gross margin over 55% in the first half, leading to an increased annual target. IDC data shows it holds the top market share in China's enterprise MaaS sector.
Source: 晚点LatePost
Read on 88z.aiLibTV has launched a Double 11 promotion for AI video creation, offering unlimited zero-point usage of its three flagship models, Wan 3, Kling 4.0, and Minimax H3, for up to 30 days. Seedance 2.5 720P is priced as low as 0.17 yuan per second. Annual members can save 1300 yuan with red packets, receive an additional three months of membership and over 170,000 points, while the Supreme 92K annual package sees a significant price reduction. The creator growth program has also been upgraded, with in
Source: LibTV
Read on 88z.aiGoogle DeepMind has released its flagship model, Gemini 4 Carbon, which topped AI intelligence benchmarks and surpassed key competitors. The model features long-term processing capabilities for large codebases, focusing on security vulnerability detection and repair, and is currently available only to select cybersecurity professionals.
Source: APPSO
Read on 88z.aiModelBest open-sources the MiniCPM-OCR series, with the 1.3B model achieving SOTA on two benchmarks and the 2B model being the industry's first OCR result discrimination model.
Source: 面壁智能
Read on 88z.aiTencent has launched Buddy One, a cross-app AI collaboration system for individual users. It integrates Tencent Docs, Meeting, and ima by sharing the WorkBuddy Agent foundation and unifying accounts and point pools.
Source: 腾讯WorkBuddy
Read on 88z.aiGoogle DeepMind has released Genie 4, a model that generates interactive virtual worlds from text, featuring real-time physics simulation and native audio generation for synchronized sound effects and character dialogue.
Source: 量子位
Read on 88z.aiAnt Group has released Lingguang 2.0, an upgraded personal AI assistant with cross-session memory and long-term task scheduling capabilities. Relying on independent cloud hosts, it transitions from an executor to a manager of affairs, proactively handling finance, childcare, chores, and travel. The system integrates the APASS security framework, using separated hosting for authorization keys and Agents to ensure privacy, allowing users to adjust or revoke permissions at any time.
Source: 蚂蚁集团
Read on 88z.aiBaidu Wenxin has released and open-sourced Unlimited OCR 2, an end-to-end document parsing model capable of continuously recognizing documents exceeding 3,600 pages. Its core SR-SWA attention mechanism, combined with chunked decoding, reduces peak KV cache for hundred-page tests by approximately 96% and increases 16-way concurrent throughput by 3.65 times.
Source: 量子位
Read on 88z.aiMicrosoft has released the MAI-Image-2.7-Preview image editing model. AI CEO Suleyman stated that it ranks second globally on the Artificial Analysis leaderboard, trailing only OpenAI's GPT Image 2.5 Sunburst and surpassing Google, Meta, and xAI. As a successor to the MAI-Image-2.6 released in August, this model scores 24 Elo points higher than the next best model in scene and style editing.
Source: IT之家
Read on 88z.aiAnthropic has publicly filed for a Nasdaq IPO under the ticker "ANT," targeting a valuation of up to $2 trillion. Although H1 revenue exceeded $16 billion, significant paper losses were driven by changes in the fair value of financing instruments. Growth is fueled by Claude Code and enterprise clients, while massive long-term compute contracts bring challenges related to high capital expenditure and supply chain dependence.
Source: 量子位
Read on 88z.aiSpaceX CEO Elon Musk announced that the company has successfully completed $40 billion in debt financing, primarily to purchase GPU chips, with delivery expected in the first or second quarter of next year. Capital expenditures for the fourth quarter are projected between $35 billion and $40 billion. Musk noted that companies lacking revenue sources outside AI data centers will face financing difficulties, potentially leading to consolidation or bankruptcies among new cloud providers, though Spa
Source: 财联社
Read on 88z.aiOpenAI has released GPT-6.1 Terra, focusing on low cost, high speed, and AI Agent office scenarios. The model features significant enhancements in visual understanding, computer operation, factual accuracy, and coding, capable of handling blueprints, invoices, desktop software, and research tasks. Its standard API price is 50% lower than the previous generation, and it has been integrated into ChatGPT Work and Codex.
Source: APPSO
Read on 88z.aiJimeng AI has launched a Short Drama Studio, integrating self-developed models to cover the entire workflow from script generation and art design to storyboard preview and post-production. It supports long-text analysis, 3D director preview, and voice reconstruction, while facilitating global localization through remake-style image translation.
Source: 即梦AI
Read on 88z.aiTencent Hunyuan has open-sourced Hy-SimulMT, a simultaneous interpretation model supporting real-time Chinese-English translation and voice cloning. With a response time of about 3 seconds, it is faster than human interpreters and achieves expert-level quality, outperforming mainstream open-source models and many commercial APIs.
Source: 腾讯混元
Read on 88z.aiMoonshot AI has officially released Kimi K3.1 Pro, the first model in the K3.1 series, featuring 3 trillion parameters and support for a million-token context. It significantly enhances capabilities in coding, 3D creation, financial research, and long-horizon agents, while introducing a high-speed version that is approximately two to three times faster. The model is now available via the official website, App, Work, Code, and API, with full weights scheduled for release before November 4.
Source: 月之暗面 Kimi
Read on 88z.aiTencent Hunyuan has released Hy-Nomas, an open-source interactive video world generation model. It supports text and camera trajectory control while maintaining long-range consistency. Optimized for local inference on a single RTX 4090 with LoRA fine-tuning support, it is suitable for gaming, simulation, and film pre-visualization.
Source: Hy-Nomas
Read on 88z.aiVolcengine has released the premium video generation model Seedance 2.5 4K, focusing on optimizing facial expressions, subtle movements, and visual details. The model supports 10-bit color depth, significantly enhancing texture quality, color layers, and gradient effects. It also maintains better clarity and consistency of materials during reference generation and video editing tasks. The workflow allows users to generate 480P drafts for storyboard verification before producing the final 4K vide
Source: 火山引擎
Read on 88z.aiMiniMax has open-sourced the weights for its MiniMax M3.1 Flash model. Featuring 428 billion total parameters and 23 billion activated parameters, it supports a context window of one million tokens. Within 72 hours of launch, it processed over 100 trillion tokens, with inference costs cut by more than half compared to the previous generation. It ranks first among domestic models in several Agent benchmarks. Ten chip manufacturers, including Huawei Ascend and NVIDIA, along with inference framewor
Source: MiniMax 稀宇科技
Read on 88z.aiiFlytek has launched the edge-cloud integrated Spark X2.6, SparkAudio 1.0 voice foundation model, and embodied multimodal interaction technology. The company also introduced AI glasses, Robot Super Brain Platform V3.0, and other software and hardware products. Emphasizing full-stack domestic computing power and edge-cloud synergy, iFlytek focuses on security, privacy, and cost-effectiveness, with applications in education, healthcare, law, and the developer ecosystem.
Source: 科大讯飞
Read on 88z.aiTencent Hunyuan and Fudan University jointly launched the WorkForge framework, designed to scale the construction of verifiable agent work environments from real-world resources. They simultaneously released the WorkForge-16K dataset, covering 16,700 work environments across 40 professional fields. The WorkForge-27B model, trained on real workspace trajectories from this framework, performed excellently, comprehensively surpassing models with hundreds of times more parameters in five major work
Source: 腾讯混元
Read on 88z.aiOpenAI has partnered with Google Cloud to bring its advanced models and Codex agent workflows to Google Cloud Model Garden, with a launch expected within weeks. Enterprise customers can count OpenAI model costs toward their GCP committed spend while leveraging Google Cloud's data governance and high-performance infrastructure. OpenAI models will also be integrated into Google Antigravity and Gemini Agent in the future.
Source: OpenAI
Read on 88z.aiThe full version of Kling 4.0 has launched on LibTV, supporting 30-second native video generation, enhanced physical consistency and director-level control, along with upgraded audio-visual performance.
Source: 哩布哩布AI
Read on 88z.aiAnthropic has launched Claude Duet, an AI agent for SMEs and professionals built on the Opus 5.5 model. It features daily briefings, app monitoring, dashboard creation, meeting scheduling, and financial forecasting, with an independent browser VM for background operation, directly competing with OpenAI Dots and Meta Muse.
Source: 财联社
Read on 88z.aiThe Memory Tensor team has open-sourced the MemT 1.0 series of foundation models, including Flash and Nano versions. This series integrates updatable internal memory into computation, allowing access to historical information without retrieving original text. In 11 memory benchmarks, the two models achieved average scores of 50.6 and 47.0, ranking top two among parameterized memory methods. Technical highlights include Gated DeltaNet-2 per-token updates, MoS hybrid storage expansion, and a two-s
Source: 记忆张量MemTensor
Read on 88z.aiOtter.ai announced a $320 million non-dilutive financing round from General Catalyst's Customer Value Fund, aimed at driving business growth, market expansion, and product innovation. The company had previously raised a total of $50 million, with last year's annual recurring revenue exceeding $100 million and its user base growing from 25 million to 40 million.
Source: Z Potential
Read on 88z.aiXingchen Lab has open-sourced the intent understanding model Xing4.0-29B-Intent, releasing 2.75 million training corpus entries, model weights, and inference images. The model uses a unified semantic parsing framework to support intent classification and parameter extraction, converting natural language into structured JSON. Evaluations show a comprehensive score of 58, surpassing the top score of 49 achieved by 18 mainstream general large models.
Source: 中国电信天翼AI
Read on 88z.aiDeepSeek has announced the full open-sourcing of its internal inference framework, DeepInfer. Integrating high-performance operators and an optimization engine, the framework supports distributed inference and advanced decoding techniques. It is compatible with NVIDIA and Ascend hardware and currently powers the high-throughput, low-latency services of DeepSeek V4.1 Flash.
Source: DeepSeek
Read on 88z.aiCleanYourTV is an open-source Android TV cleanup tool that requires no root access. It connects via ADB to automatically disable ads, telemetry, and bloatware using AI assistants, replacing them with a lightweight desktop. With extensive device support and one-click restore, it has recently gained significant popularity.
Source: CleanYourTV
Read on 88z.aiAI assistant Today announced that its cumulative funding has surpassed $200 million, with investment from IDG Capital, Tencent, Wuyuan Capital, Ant Group, Sequoia China, and Transsion Holdings. The product aims to build a personalized proactive AI system covering schedule management, life assistance, and work support.
Source: Today AI 助理
Read on 88z.aiMiniMax has released the new generation model MiniMax M3.1 Flash, achieving generational improvements in agentic tasks, programming, and cybersecurity. It offers approximately three times faster inference speed and a seventy percent reduction in comprehensive costs, with top domestic rankings in multiple benchmarks.
Source: MiniMax 稀宇科技
Read on 88z.aiTencent has launched 'Dongle Ya', an AI learning community for parents and teachers. It uses 'Wonder Cards' to turn picture books and animations into interactive content. With a built-in AI Agent, users can generate multimodal materials via voice, then adapt, publish, and share them on WeChat and QQ, covering K12 classrooms and extracurricular exploration.
Source: 懂了鸭
Read on 88z.aifal Research and Black Forest Labs have jointly released FLUX 3 Max, a post-trained video model supporting text-to-video, image-to-video, and keyframe interpolation. It generates 5 to 20-second videos in 720p or 1080p with native audio, emphasizing extreme speed and claiming to be 22 times faster than Seedance 2.5 at similar quality.
Source: bfl.ai
Read on 88z.aiAnt Bailin has open-sourced the Ming-3.0 voice model series, comprising five models covering speech synthesis, audio understanding and dialogue, and speech recognition. These models can be flexibly combined for business workflows. The TTS model supports emotion and rhythm adjustment with paralinguistic tags, the audio model provides structured summaries of long recordings, and the ASR model supports hotword injection with an average error rate of 4.41%.
Source: 百灵大模型
Read on 88z.aiDeepSeek releases DeepSeek-V4.1-Flash-1017, updating web, app, and API models. The new version improves baseline performance, reduces latency and output length, optimizes specific frameworks, and enhances wish-based programming capabilities.
Source: DeepSeek
Read on 88z.aiTencent Hunyuan has released Hy Memory 2.0, an open-source agent memory framework featuring specialized models, the framework itself, and the MemHub application. Covering chat, work, and code memories, it ranks first among open-source solutions in four benchmarks. The core Hy-Memory-30A3B model has approximately 3 billion activated parameters and outperforms Opus5 in average scores. MemHub connects with ten AI tools like Cursor for real-time memory synchronization and is already deployed in Tenc
Source: 腾讯混元
Read on 88z.aiLovart has officially released version 2.0, introducing a personal design agent named "Arti." This assistant builds a memory bank to maintain user aesthetic preferences, proactively conducts design research, and generates trend reports and mood boards. It monitors canvases in the background to ensure style consistency, handles annotations and comments, and accumulates reusable skills. The new design commission feature supports 24-hour automated execution of repetitive tasks such as competitor mo
Source: APPSO
Read on 88z.aiKling AI announces that Kling 4.0 is now fully available on its official website, with upcoming support for 4K and 1080p 10-bit HDR output, a comprehensive upgrade to the creation page, integration of capabilities into MCP/CLI, and imminent opening of enterprise API services.
Source: 可灵AI
Read on 88z.aiBlack Forest Labs has open-sourced the FLUX 3 Image Klein model, featuring 5.6 billion parameters and enabling local operation on a single consumer-grade GPU. Built on FLUX 3's cross-modal shared learning, it offers generation, editing, and multi-reference capabilities, with precise layout control via bounding boxes and strong performance in text rendering and realistic details.
Source: bfl.ai
Read on 88z.aiMeta has released Muse Spark, an open-source native multimodal model featuring a mixture-of-experts architecture with 816 billion total parameters and 42 billion active ones. Supporting a million-token context, it targets coding and agents. While topping six benchmarks and claiming to be the strongest US open-weight model, it trails Chinese counterparts like Kimi, Qwen, and Zhipu in software development.
Source: 财联社
Read on 88z.aiOpenAI has released GPT-Live-1.1, a new voice model for ChatGPT, with its lightweight Quip version replacing mini. The model supports full-duplex interaction, adds six new languages and the English Bubbie voice, and offers more natural and diverse search feedback.
Source: GPT-Live-1.1
Read on 88z.aiSpaceXSI has released Grok Voice 3.0, claiming it is the world's most intelligent speech-to-speech model with latency under one second. The model ranks first on the Artificial Analysis comprehensive leaderboard, achieving a 97% score in full-duplex benchmarks and 71.6% in τ-Voice. Its performance is comparable to GPT-Live-1 but with 40% lower latency. Currently, the model supports millions of daily conversations in Tesla vehicles and handles 20,000 calls per day at Starlink customer service cent
Source: IT之家
Read on 88z.aiKunlun Wanwei has open-sourced Matrix-Game 4.0, the world's first video large model game capable of real-time operation on a single consumer-grade GPU. Running on an RTX 5090, it achieves 480p resolution at 60 FPS with end-to-end latency under 150ms and peak VRAM usage of 18.5 GiB. Its architecture involves a game engine maintaining world state, a Coding Agent generating or rewriting rules, and a video model handling real-time rendering.
Source: 昆仑万维集团
Read on 88z.aiStepFun has officially released the weights for its Step 5 Preview model, which are now available on Hugging Face and ModelScope. Utilizing a sparse MoE architecture, the model features 600 billion total parameters with only 27 billion activated. It supports a context window of one million tokens and natively handles both text and visual inputs. The model demonstrates strong performance in long-horizon agent tasks such as AI coding and finance, ranking second globally in 第三方模型接口 call volume.
Source: 阶跃StepFun
Read on 88z.aiTencent has launched 'Rena Lancai', a personal stock analysis Agent for ordinary investors, accessible via web, WeChat Mini Programs, and WeChat Assistant. The tool automatically links daily quick reads, market briefs, and watchlist updates. It supports importing holdings, viewing individual stock dynamics, and providing reasons and verification standards for bullish, bearish, or neutral judgments, allowing users to take notes and ask follow-up questions.
Source: Rena览财
Read on 88z.aiGoogle's new flagship model, Gemini 4 Argon, is now available to global developers and enterprise users. Access is provided through the Gemini API, paid tiers on the Gemini platform, Google Antigravity Ultra membership, or Ultra and Pro memberships for the Gemini app and search. The model has undergone limited testing and extensive internal real-world validation, focusing on complex multi-step tasks in programming, scientific research, law, and finance.
Source: APPSO
Read on 88z.aiModelBest and OpenBMB have jointly released the open-source edge multimodal model MiniCPM-V 4.7, available in 35B-A3B and 1B versions. The 35B-A3B is the world's first open-source MoE multimodal large model capable of lossless, smooth operation on edge devices, requiring only about 3GB of memory and performing well in various benchmarks. The model supports multiple chips and its code is fully open-sourced.
Source: 面壁智能
Read on 88z.aiAlibaba's Qwen Office introduces an e-commerce version, an AI business assistant designed for merchants. It connects multiple platforms via a unified entry point, simplifying login, report downloading, and verification processes, while supporting global business overview, trend capturing, ROI optimization, and bulk content generation.
Source: 阿里巴巴
Read on 88z.aiAnthropic has released Scion, a self-developed version control system designed to replace Git, which has been in use for two decades. Addressing synchronization bottlenecks and massive traffic pressures caused by the surge in Claude-assisted coding, Scion adopts an always-online centralized architecture. It significantly optimizes code submission and retrieval efficiency through virtual checkouts and streaming pull technologies.
Source: APPSO
Read on 88z.aiAlibaba has launched QwenGyre, an online reinforcement learning framework designed for ultra-long-horizon agent tasks. The framework reduces GPU idling through elastic scheduling and controls trajectory processing to lower the cost of forked training. Tests on the Qwen3.8-2.4T model show a significant increase in NL2RepoBench pass rates for single rollouts of approximately 700k tokens. Under the same GPU budget, it achieves faster acceleration and greater training time savings compared to Async,
Source: 千问大模型
Read on 88z.aiMeta has released the new Muse Wonder model and its enhanced version, Wonder Full. These models perform close to Fable 5.1, Opus 5.5, and GPT-6 Astra in various benchmarks. Focusing on active memory and sustained attention, the series aims to realize the vision of personal superintelligence, with specific optimizations for cross-conversation memory, long-process tasks, and coding performance. Zuckerberg emphasized accelerating AI investment and advocating for industry self-regulation to ensure s
Source: 财联社
Read on 88z.aiZhipu has released GLM-Legal, a legal-specific large model built on the GLM-5.3-Flash base. It covers the entire workflow from contract review to document generation. The model achieved a 19.2% pass rate in Harvey LAB tests, matching leading closed-source models, and supports API and AutoClaw desktop access with a focus on low cost and traceable evidence.
Source: GLM-Legal
Read on 88z.aiCharacter.AI has filed for Chapter 11 bankruptcy protection in the US and agreed to sell certain intellectual property assets to Disney for $20 million. Disney serves as the stalking horse bidder with its offer setting the floor price, allowing others to bid higher. The deal is expected to close between late November and early December. Operations continue during the sale process, but the platform will gradually shut down if no buyer is found.
Source: IT之家
Read on 88z.aiElon Musk's SpaceXAI official account has been officially renamed to "SpaceX Super Intelligence" with the new handle @SpaceXSI. This move confirms Musk's previous statement positioning SpaceX as a super intelligence company, marking a comprehensive brand transformation of its AI business toward super intelligence.
Source: 财联社
Read on 88z.aiHeyGen has released its proprietary speech model, HeyGen Voice, which topped the Artificial Analysis TTS leaderboard. Sharing the same end-to-end tech stack as the Avatar V digital human model, it preserves user tone, stress, emotion, and rhythm to improve reading naturalness. The decoupling of similarity and expressiveness_boost parameters balances fidelity and expressiveness.
Source: Voice
Read on 88z.aiInfinigence AI announced the completion of a 2.3 billion yuan Series C funding round, bringing its total fundraising to 4.3 billion yuan, setting new records in the domestic AI infrastructure sector. Investors include state-owned capital, industrial partners, and prominent firms such as Sequoia China, Tencent, and Baidu. The company focuses on efficiency optimization between chips and models without building data centers or training models. Its platform supports 16 chip types and 37,000 PetaFLOP
Source: 量子位
Read on 88z.aiMusi Intelligence announced the completion of its Pre-A financing round, raising over 1 billion yuan. The round was led by Huakong Fund, with participation from Guokai Sci-Tech Innovation, Shanghai Sci-Tech Innovation Fund, and existing shareholders. The funds will be dedicated to researching and developing full-modal real-time interactive intelligence and advancing global commercialization. Incubated by Qiu Xipeng's team, the company aims to extend AI beyond screens into devices like glasses, c
Source: 模思智能
Read on 88z.aiByteDance announces the merger of TraeWork and TraeCode into the new TRAE, upgraded to a full-link development platform supporting multi-terminal synchronization. The new platform offers a unified entry point, allowing seamless switching between Agent mode for end-to-end delivery and IDE mode for deep coding, covering desktop, web, and mobile.
Source: TRAE.ai
Read on 88z.aiTencent Yuanbao has launched 'Yuanbao Desktop Pet', a desktop companion featuring a green mascot that provides task progress updates, idle companionship, and lifestyle reminders.
Source: 元宝
Read on 88z.aiTypeSafe AI, developer of the decision model Jev, has announced the completion of an $870 million funding round, reaching a valuation of $7.5 billion. Since its launch, the model has gained rapid adoption, being used by approximately one-third of Fortune 500 companies, surpassing one million users, and processing trillions of tokens daily.
Source: IT之家
Read on 88z.aiOpenAI has released the Ultrafast version of GPT-6.1 Sol, delivering an 8x speed increase while maintaining intelligence levels close to the flagship Astra model. Designed for real-time interaction scenarios, it is priced at $12 per million input tokens and $60 for output, with access currently limited to specific subscribers and enterprise or education clients.
Source: 机器之心
Read on 88z.aiGoogle has launched the general office agent Gemini Agent, featuring autonomous planning and cross-application collaboration capabilities. This agent integrates into teams with an independent identity and optimizes task execution through four memory mechanisms, automatically selecting underlying models based on requirements.
Source: 量子位
Read on 88z.aiJetBrains has released the open-source model Mellum2.1, featuring a 12B mixture-of-experts architecture. The model makes reinforcement learning central to training, supplements data across multiple domains, and launches millions of sandboxes. It offers capabilities for codebase exploration, file editing, and fault diagnosis, with inference throughput nearly double that of Qwen3.5-9B under high load and a speed increase of approximately 1.6x via multi-token prediction.
Source: IT之家
Read on 88z.aiAI model evaluation platform Arena announced the completion of a $200 million Series B financing round, reaching a valuation of $3.1 billion. The round was led by Lightspeed and Khosla, with participation from a16z, Salesforce, and others. Since its Series A in January, its valuation has nearly doubled within ten months, while annual recurring revenue grew from $30 million to $100 million.
Source: Z Potential
Read on 88z.aiGoogle DeepMind has released Nano Banana 2.1, an image model built on Gemini 3.6 Flash. The update focuses on improving design completeness, local editing, multi-subject consistency, and visual naturalness, with benchmark scores showing significant gains in infographic design and character consistency.
Source: 机器之心
Read on 88z.aiGoogle has released EmbeddingGemma 2, built on the Gemma 4 architecture with 740 million total parameters. The model maps text, code, images, video, and audio into a unified embedding space. Its modular design requires only 270 million parameters for text-only tasks, allowing it to run on the Pixel 11 Pro with approximately 191MB of memory after quantization.
Source: IT之家
Read on 88z.aiButterfly Effect, the parent company of Manus, has completed over $500 million in financing, led by Boyu Capital and IDG Capital, with participation from existing shareholders including Tencent. This marks its first funding disclosure since resuming independent operations. Meanwhile, Manus released version 2.0, introducing its self-developed Cascade framework, cloud computers, automated workflows, and the personal intelligent assistant Cue.
Source: APPSO
Read on 88z.aiAnthropic has launched its latest lightweight model, Claude Haiku 5.5, which outperforms DeepSeek V4.1 Flash and GLM-5.3-Flash in benchmarks. Positioned as a competitor to GPT-6 Luna, it offers a lower price point, with a 90% reduction in costs for inputs and outputs within 100,000 tokens. The model is the first to support five levels of effort adjustment, achieving a 72.4% score in the Max setting of the OSWorld 2.1 computer operation test. While powerful, Haiku 5.5 does not replace Sonnet 5.5
Source: 量子位
Read on 88z.aiShengshu Technology has released Vidu Q4 Preview, a new flagship video generation model that significantly enhances character performance and camera dynamics. It supports multiple reference inputs and 2K/4K high-definition output, covering the entire workflow from creative testing to final delivery.
Source: 生数科技
Read on 88z.aiMistral has released its flagship model, Mistral Large 4, featuring 1 trillion parameters and a mixture-of-experts architecture. In third-party evaluations, its intelligence index matches GPT-6 Luna, outperforming GPT-6 and Claude in several rigorous tests, with an 82% cybersecurity vulnerability fix rate. Trained on 3,800 Blackwell chips and supporting over 160 languages, the model's weights will be fully open-sourced by the end of the month.
Read on 88z.aiGoogle has launched its flagship model Gemini 4 Argon, which performs excellently in multiple benchmarks and supports million-token long outputs. It is currently available only to vetted cybersecurity defense teams.
Source: 量子位
Read on 88z.aiBilibili's Index LLM team has released the Index-Translate multilingual translation model family, featuring preview weights for 2B, 9B, and 35B-A3B parameter sizes. Built on the Qwen3.5 architecture, the models support translation across 150 languages and have been open-sourced on Hugging Face and ModelScope. Beyond basic instruction following for terminology and format retention, the models are optimized for speech subtitle and dubbing scenarios, allowing translation length adjustment by target
Source: IT之家
Read on 88z.aiAnt Group's Bailin team has released the new large model Ling-3.1-flash, featuring approximately 560B total parameters and 25B activated parameters, with a 1M context window. Using a hybrid architecture, the model focuses on real-world scenarios such as office work, healthcare, and software development, demonstrating the ability to handle complex tasks like building compilers from scratch.
Source: 百灵大模型
Read on 88z.aiDeepSeek has officially open-sourced infrastructure components for the Huawei Ascend platform, including TileLang, compute libraries, and distributed communication libraries, corresponding to NVIDIA platform components. TileLang supports DeepSeek V4 series model training, while other components cover core capabilities like matrix operations, with performance nearing hardware limits.
Source: DeepSeek
Read on 88z.aiOpenAI released 25 updates at DevDay 2026 in San Francisco, introducing the persistent agent Dots and the cost-effective model GPT-6.1 Sol.
Source: 量子位
Read on 88z.aiLibTV has partnered with the Douyin AI Creation Wave Plan to launch the 'AI Master Public Course'. Featuring Liu Shuheng and Wang Yuchen, creators of the award-winning short film 'He Long', the course deconstructs AI short film creation from narrative, art, and performance perspectives. Focusing on underlying creative logic rather than prompt engineering, it is available on Douyin.
Source: 哩布哩布AI
Read on 88z.aiAMD announces an all-stock acquisition of World Labs, with Fei-Fei Li joining as an executive.
Source: APPSO
Read on 88z.aiKuaishou's Kling 4.0 will officially launch in October, with the lightweight version already available to Black Gold annual members. The new version significantly upgrades realism, controllability, and narrative integrity, supporting up to 30 seconds of native video generation, multi-keyframe control, and precise editing. Audio is upgraded to dual-channel stereo, with upcoming 4K HDR output and two-minute extended shooting features.
Source: 可灵AI
Read on 88z.aiAnthropic has officially released Claude Sonnet 5.5, the second model in the Claude 5.5 series, positioned to complement Opus 5.5. The new model offers over 30% faster speed, significantly enhanced coding capabilities, and a Terminal-Bench 4.0 score jump from 10.3% to 70.6%. Its performance nears that of Opus 5.5 at half the price.
Source: 机器之心
Read on 88z.aiManus releases the re-architected Manus 2.0, leveraging its self-developed Cascade framework for improved performance and reduced costs, alongside Cue, a personal agent app with independent digital identities.
Source: APPSO
Read on 88z.aiThe official DeepSeek Harness desktop client has been revealed, allowing users to assign tasks to Agents directly after logging in via account or API Key and selecting a local workspace, eliminating the need for command-line startup. It offers four modes including Standard and PTC, with six built-in plugins such as Agent Team and voice input, maintaining its extensible architecture.
Source: 机器之心
Read on 88z.aiMeituan has released the new LongCat-2.5-Preview large model, maintaining a MoE architecture with 1.6T total parameters and approximately 48B activated parameters. It supports a 1M token ultra-long context window and a maximum output of 128K tokens. The model natively integrates image understanding capabilities, enabling cross-modal Q&A and visual reasoning, and can execute long-process Agent tasks for terminals, browsers, and desktop software.
Source: LongCat-2.5-Preview
Read on 88z.aiMiniMax has released its latest text coding model, M3.1-Flash-Preview, integrated into the MiniMax Code platform. Designed for daily development, it supports workflows from bug fixing to full feature delivery, with enhanced capabilities in requirement understanding, boundary handling, and testing verification. Additionally, the company has reset all user Token allowances and introduced free claim benefits.
Source: IT之家
Read on 88z.aiKuaishou has launched 'likli', an AI video creation agent supporting macOS and Windows desktops. It covers the entire workflow from understanding requirements and writing scripts to generating shots, dubbing, and editing, delivering ready-to-use videos. The product features conversational interaction with visible and intervenable processes. It includes built-in skills such as playwright, voice actor, and product director, suitable for e-commerce, short dramas, brand ads, and spoken content.
Source: likli
Read on 88z.aiZhipu released a Trust.patch announcement, updating the open-source version of ZCode to v3.14.3. The repository snapshot upload link has been removed, and involved cloud data has been deleted. Future releases will sync with the official version. Paying users and returning users within one month receive multiple reset card benefits.
Source: ZCode
Read on 88z.aiNetEase Youdao has open-sourced two real-time interaction models, R2T2 and T3PO, from its Ziya Live series. Optimized for streaming speech recognition and translation respectively, they quickly topped relevant trending lists on Hugging Face.
Source: 网易有道
Read on 88z.aiTencent Music has launched MusicBuddy, an AI music creation platform offering end-to-end services for independent musicians from inspiration to management. The platform retains multi-track editing capabilities, allows arrangement and mixing via natural language commands, provides data-driven creative suggestions, and automatically generates promotional materials.
Source: MusicBuddy
Read on 88z.aiStepFun has released Step Code, an open-source terminal coding agent under the MIT License. This tool handles coding, debugging, execution, and delivery directly within the terminal. It tied for first place in Terminal Bench 2.1 with an 80.9% pass rate while using fewer tokens. In the self-built Multi-Frame long-horizon benchmark, Step Code led six Agent Harness models with a 73.3% pass rate and the lowest average token usage of 5.09M.
Source: 阶跃星辰开放平台
Read on 88z.aiAlibaba Tongyi Qwen has launched the Qwen-Audio-3.1 series, featuring five voice models that cover the full spectrum from understanding to creation. The series enhances multilingual and dialect recognition and transcription, improves comprehension of emotions and environmental sounds, and achieves cross-language natural timbre synthesis along with unified generation of vocals and sound effects. The real-time interaction model supports full-duplex empathetic dialogue, language switching, and Agen
Source: 阿里语音AI
Read on 88z.aiAnt Group's inclusionAI has open-sourced the Ming-Image-0.1-Design series, featuring two 6-billion parameter models. The Design model generates complete UI and poster designs from text prompts, while the Layer model splits designs into 2 to 9 independently editable transparent layers. The series achieved top rankings in authoritative benchmarks and demonstrated significantly faster inference speeds compared to larger models.
Source: 百灵大模型
Read on 88z.aiChina Telecom Xingchen Lab has open-sourced TeleOCR, a document parsing model using a lightweight vision-language architecture with approximately 1.2 billion parameters. It uniformly processes digital and photographed documents, converting text, layout, tables, formulas, and scientific charts into structured results. The model topped the OmniDocBench v1.6 leaderboard with a total score of 96.87, surpassing competitors like MinerU2.5-Pro, and also secured first place on PureDocBench and won the I
Source: 中国电信天翼AI
Read on 88z.aiTencent announced that due to business adjustments, the QClaw WeChat remote work AI assistant will officially cease operations on December 24, 2026. As of now, new user registrations and subscription renewals have been suspended. Existing users' service benefits remain valid until expiration, and refunds can be requested for unused points. The official recommendation is for users to back up data locally or migrate it to the WorkBuddy platform via the official website.
Source: IT之家
Read on 88z.aiOpenAI has released the GPT-6 Sol and Luna models, cutting API prices by half compared to the 5.6 series. Luna's input price is lower than DeepSeek's off-peak rate, achieving GPT-5.6 Sol performance at about one-tenth the cost. Sol approaches Astra's performance with improved caching and inference efficiency, at a much lower cost than Claude Opus 5 and Fable 5.1.
Source: 量子位
Read on 88z.aiAnthropic has released Claude Opus 5.5, the first model in its Claude 5.5 series, setting new state-of-the-art records in agent coding, computer use, and knowledge work. It scored 66.4% on Terminal-Bench 4.0, outperforming GPT-6 Astra by 8.5 percentage points, and achieved 1846 Elo on the GDPval benchmark. Additionally, the model is 40% cheaper, 30% faster, and features significantly improved writing and communication capabilities.
Source: APPSO
Read on 88z.aiAishix Technology has launched PixVerse R2, the world's first general real-time world model with scaling capabilities. It supports applications in gaming, interactive film, and digital humans. By leveraging Omni Causal AR expansion and real-time acceleration, the model balances real-time performance with generality while enhancing quality, long-range consistency, and multimodal control.
Source: 爱诗科技 AIsphere
Read on 88z.aiAlibaba has released Qwen Intelligence, a non-hardware full-stack AI solution for phone manufacturers. It includes scenario-optimized models, a customizable platform, and three core agents, significantly improving task execution efficiency and generation speed.
Source: 千问大模型
Read on 88z.aiAiPy has officially launched on the KylinOS Software Store with one-click installation. Combining large models with a Python environment, it executes tasks like data analysis and document generation via voice or natural language commands. It supports multiple platforms, local intranet deployment, and domestic large models, featuring enterprise-grade security capabilities.
Source: 爱派AiPy
Read on 88z.aiMoonshot AI has officially launched the Kimi Code desktop client, with simultaneous support for macOS and Windows. The application brings AI agent capabilities to a graphical interface, enabling developers to handle requirement expression, progress tracking, debugging, and Git collaboration within a single window. This reduces the barrier to entry and suits various scenarios ranging from prototype development to complex engineering refactoring.
Source: 月之暗面 Kimi
Read on 88z.aiSpaceXAI has officially released its flagship model, Grok 4.7, featuring 2.1 trillion parameters, a roughly 40% increase from the previous generation. The model shows enhanced capabilities in programming and knowledge work, outperforming GPT-5.6 Sol Max in several benchmarks. However, it trails behind Claude Fable 5.1 and OpenAI GPT-6 Astra. While lagging in specific tests like Terminal-Bench, it ranks first in the second tier for comprehensive intelligence index.
Source: 机器之心
Read on 88z.aiXiaomi has released and open-sourced the Xiaomi MiMo-V2.6 series, featuring two native full-modal models: Pro and Flash. After extensive reinforcement learning training, the Pro version surpasses several well-known models in comprehensive intelligence index, becoming the strongest open-source model, though it still lags behind top closed-source counterparts.
Source: Xiaomi MiMo
Read on 88z.aiTencent Hunyuan has released the Hy Image3.5 preview model, featuring a roughly 30% performance improvement, support for 2K resolution and multi-turn editing, with an API price of 0.15 yuan per image, already integrated into apps like Yuanbao.
Source: 腾讯混元
Read on 88z.aiHiDream.ai has announced a brand refresh to 'Zhi Xiang Yuan Yi,' marking its development origin. As an early player in multimodal generative AI, the company has undergone four generations of architectural innovation, establishing a native full-modal large model matrix and a '1+1+X' business system. The new brand blends Eastern philosophy with modern computing, using 'Yuan Yi' to symbolize origins, while conveying an open temperament through minimalist visuals and a four-color system to empower c
Source: HiDream智象元一
Read on 88z.aiTencent has launched the 'Hunyuan Think Tank' platform to recruit senior external experts across more than ten vertical fields, including finance, law, and healthcare. Certified experts, vetted through online qualification reviews, will participate in high-quality dataset production, content creation, and knowledge auditing via remote tasks. Compensation is linked to delivery quality, with a dynamic rating system based on accuracy and performance, allowing higher-rated experts to prioritize high
Source: 混元智囊团
Read on 88z.aiAlibaba's Tongyi team has released the open-source image model Qwen-Image-2.1. With only 7 billion parameters in its visual generation component, it balances output quality with inference costs. The model integrates text-to-image and image editing capabilities, supporting transparent image generation and local edits, while significantly improving fidelity for portraits and products.
Source: 千问大模型
Read on 88z.aiZhipu's ZCode is now open source, promising not to retain community code data for training and establishing a security feedback mechanism. Version v3.14.0 has removed the Repo Wiki feature and related generation pipelines.
Source: ZCode
Read on 88z.aiAI film and television creation platform LibTV has released version 1.5, introducing five exclusive features that cover the entire workflow from character design to final editing. New capabilities include visual character adjustment, intelligent screenwriting assistance, director avatar for automatic shot splitting, modeling-free virtual studios, and automated post-production refinement tools, aiming to enhance creative efficiency and quality.
Source: LibTV
Read on 88z.aiCapCut PC launches CapCut Hub, integrating storyboard scripts, asset management, and batch generation capabilities. It supports one-click entry into multi-track editing and AI smart editing features. The CapCut Assistant Agent is also released, enabling voiceover editing and subtitle correction via text commands, with compatibility for importing materials from Jimeng and Xiao Yunque.
Source: 量子位
Read on 88z.aiMiniMax has open-sourced version v0.4.12 of MiniMax Code CLI, with the source code now available on GitHub. As a core component of the MiniMax Code client, the tool achieved a 76.7% task pass rate in the FrontierHarness Eval benchmark, with a median time of 4 minutes and 33 seconds for successful tasks, outperforming public baselines. The company states that this open-source release allows developers to inspect tool invocation and permission handling, fostering the joint creation of secure and r
Source: MiniMax 稀宇科技
Read on 88z.aiAlibaba Tongyi has released Qwen3.8-LiveTranslate, a real-time simultaneous interpretation model supporting 60 languages with reduced latency of 2.3 seconds per character. Utilizing a Hybrid MoE architecture, it features real-time speaker diarization and long-context disambiguation, outperforming mainstream systems in multilingual evaluations.
Source: 千问大模型
Read on 88z.aiStepFun has released its new flagship base model, Step 5 Preview, featuring a sparse MoE architecture, million-token context support, and native multimodality. It ranks among the top three global open-source models in intelligence index, offers significantly lower costs per task than competitors, and excels in long-duration Agent tasks and various professional fields.
Source: 阶跃StepFun
Read on 88z.aiAlibaba's Qoder Agent Workbench has officially launched Qoder Sites, enabling users to build and publish web applications via natural language commands. The feature automatically generates shareable links with permission management capabilities. It supports synchronous backend database creation for data persistence and state updates. Additionally, Qoder Sites allows integration with Qoder Cloud Agents to empower applications with agent capabilities, suitable for scenarios such as product prototy
Source: 阿里云
Read on 88z.aiThe HappyOyster 1.0 series models are now live on Alibaba Cloud Bailian, offering public API access without invite-only testing. The suite features two core models: Adventure and Directing. Adventure generates real-time open worlds from text or image inputs, supporting movement, camera control, and scene interaction. Directing drives character narratives and camera angles via instructions after world generation. Currently, these services are available only in the Singapore and Virginia regions.
Source: 阿里云百炼
Read on 88z.aiAlibaba Qwen releases Qwen3.8-Omni-Flash, supporting multi-modal input and 1M context, with significant performance improvements and a sharp drop in audio API pricing.
Source: 千问大模型
Read on 88z.aiZhipu has released GLM-5.3-FlashX, achieving inference speeds of up to 200 tokens/s. Previously introduced as Ox Alpha, the model is built on the computing power of 100,000 domestic chips, with enhanced infrastructure investment and inference optimization. As a leading model in intelligence for its size, the GLM-5.3-Flash series offers high cost-effectiveness, and this upgrade further strengthens its comprehensive competitiveness in intelligence, price, and speed.
Source: 智谱
Read on 88z.aiTencent's AI game creation tool, formerly codenamed 'Craft', is officially named 'Youdianling CraftBuddy'. The new name retains 'Craft' and incorporates the concept of a partner, positioning itself as the user's 'AI game creation companion'. Users can visit craft.qq.com to experience games created with Youdianling. Creation access is being rolled out gradually, and official channels should be followed for event updates.
Source: 游点灵-代号Craft
Read on 88z.aiTencent Cloud has released Octop 1.0 GA, a self-hosted multi-agent AI assistant designed for solopreneurs, families, and small teams. The update enhances knowledge retention, multi-device access, permission controls, and security. It is now available on the Lighthouse and CVM official image markets, enabling one-command deployment of a private AI collaboration hub.
Source: 腾讯云服务器
Read on 88z.aiTencent has launched Chatterfly, an AI-native voice assistant, marking its entry into the AI input sector. Positioned as an expression aid, the tool automates the process from spoken ideas to polished text. It is optimized for weak network and noisy environments and integrates with Agent products within the Tencent ecosystem.
Source: Chatterfly
Read on 88z.aiMoonshot AI's Kimi has launched a financial industry solution that integrates multiple data sources and professional skills. In partnership with China Securities Co., Ltd., it has significantly reduced report production time. The solution is now fully online and adopted by numerous institutions.
Source: 月之暗面 Kimi
Read on 88z.aiiFlytek has released Spark-Audio-1.0-Preview, a speech foundation model trained on fully domestic computing power. Utilizing a 0.65B audio encoder and a 30B-A3B MoE language model trained on 13 million hours of audio, it directly understands speech semantics, emotions, and scenarios.
Source: 科大讯飞
Read on 88z.aiShengshu Technology has released the Vidu S2 video model, featuring an Avatar version for real-time interaction and an Editing version with four editing capabilities, significantly improving image quality and dynamic consistency.
Source: 生数科技
Read on 88z.aiiFlytek Open Platform has officially launched AStudio, a desktop AI workbench. Built on the Spark X2.5 model and equipped with the ACode Harness Agent framework, it supports multi-model switching, planning modes, and multi-agent collaboration. Featuring dual spaces for daily tasks and projects, AStudio connects to local files and real-world projects to generate, deliver, and verify documents, web pages, and code, aiming to enhance AI productivity across all scenarios.
Source: 讯飞开放平台
Read on 88z.aiGoogle has released the Gemini 3.8 Live series of real-time conversation models, comprising a standard version and an Extended Thinking version. The standard edition focuses on scalable deployment and cost optimization, supporting automatic multi-language switching and visual understanding. The Extended Thinking edition enhances multi-step reasoning capabilities, enabling simultaneous reasoning and speech, and performs excellently in various benchmark tests.
Source: IT之家
Read on 88z.aiDoubao Work announces two major updates: the Doubao 2.1 Pro model has completed a minor version upgrade and is now fully available via Volcano Ark API, while the new 'Doubao Work Partner' independently integrates into Feishu for on-demand team collaboration.
Source: 豆包工作
Read on 88z.aiShanghai AI Laboratory has released ShuSheng-S2, an open-source multimodal foundation model. Its general capabilities rank among the top tier of open-source models. The model performs comparably to leading closed-source systems in scientific long-horizon tasks such as biology, materials science, chemistry, and advanced mathematics. By employing a Memory Decoder architecture with pluggable professional memory modules, ShuSheng-S2 balances domain expansion with general proficiency while strengthen
Source: 书生Intern
Read on 88z.aiUnitree has fully open-sourced its new general humanoid foundation model, UnifoLM-WLA-1.0, releasing the code, model weights, and datasets simultaneously. Trained on large-scale multimodal perception and interaction-centric world modeling using approximately 2,500 hours of real-robot data, the model achieves leading performance in embodied reasoning benchmarks, rivaling top closed-source models. It enables a single model to manage 64 tasks, covering both desktop manipulation and whole-body mobil
Source: Unitree宇树
Read on 88z.aiStepFun has released the StepAudio 3 series of voice large models, covering five areas: real-time interaction, speech recognition, speech synthesis, full audio generation, and music creation. The Realtime model topped the global real-time voice interaction rankings with a comprehensive score of 98.9%, while the ASR model achieved a word error rate of just 1.7%, tying for first place. Several models in the series have achieved leading global positions on the Artificial Analysis leaderboard, demon
Source: 阶跃StepFun
Read on 88z.aiHiDream Future has released its first native omni-modal video model, HiDream-O1-Video-1.0, which supports multi-modal inputs and generates 5-20 second 1080p videos with one click. The model features significant upgrades in intent understanding, physics simulation, and audio-visual synchronization, ranking highly on global authoritative lists and marking the entry of Chinese AI video technology into the world's top tier.
Source: HiDream智象元一
Read on 88z.aiHe Mujun, President of Sugon Intelligent Computing, revealed at the 2026 Computing Power Network Development Conference that the Supercomputing Internet platform has over 1.7 million registered users. It aggregates 3.5 million CPU cores and 250,000 GPU cards, with domestic computing power accounting for more than 99%. The platform has built six major agent matrices, launched over 400 agents and 10,000 skills, and partnered with dozens of institutions including Peking University and BGI.
Source: 超算互联网
Read on 88z.aiMoonshot AI has released its new flagship model, K2.8 Preview, fully integrated into Kimi Code and Kimi Work. Supporting a 1M context window, the model features enhanced coding and agent capabilities, with overall performance approaching the flagship Kimi K3. It is available to all membership tiers. Previously, Kimi K3's state-of-the-art frontend capabilities helped drive the company's ARR from $300 million in June to over $1 billion in August.
Source: 量子位
Read on 88z.aiZhipu announced the completion of approximately $5 billion in financing, comprising a $2 billion equity placement and $3 billion in zero-coupon convertible bonds. The funds will prioritize R&D for the next-generation GLM foundation models and computing infrastructure, with 60% allocated to self-training systems and recursive model improvement. In the first half of the year, revenue reached 954 million yuan, up nearly 400% year-on-year, with an ARR of $1.6 billion.
Source: 智东西
Read on 88z.aiAmazon Web Services has released the desktop version of Amazon Quick and its mobile smart dynamic feed feature, supporting multiple platforms and integrating key business channels while emphasizing local data residency and compliance security.
Source: 亚马逊云科技
Read on 88z.aiFollowing the open-source release of its 14B main model, Ant Lingbo Technology has launched three additional models for the LingBot-World 2.0 series. These include a Small version for consumer GPUs, a bidirectional Teacher model for high-quality distillation, and a causal pre-training base supporting multi-modal inputs, aiming to lower deployment barriers and improve training efficiency.
Source: 蚂蚁灵波科技
Read on 88z.aiOpenAI announces the API launch of GPT-Live-1, a voice model supporting full-duplex conversation and complex scenario handling, suitable for telephony agents.
Source: IT之家
Read on 88z.aiSenseTime's Xiao Huanxiong mobile app has officially launched, supporting both iOS and Android platforms. The application features core capabilities such as controlling computers via mobile commands, remotely viewing and sharing results, and long-press voice input, enabling seamless cross-device collaboration.
Source: 商汤科技SenseTime
Read on 88z.aiOpenAI has updated ChatGPT Voice Mode, allowing Pro users to select between GPT-5.6 Sol and GPT-6 Astra models and adjust reasoning intensity during voice conversations. The system automatically invokes the chosen model for search or complex reasoning tasks, with each invocation counting toward message limits.
Source: IT之家
Read on 88z.aiMoonshot AI's Kimi has launched a corporate partner program, collaborating with IT service providers and integrators to form frontline deployment teams that drive AI into core business operations. Companies including Huasheng Tiancheng and Kingsoft Cloud have signed on to co-build industry solutions.
Source: 月之暗面 Kimi
Read on 88z.aiDeepSeek has officially open-sourced DeepSeek V4.1 Flash, a Mixture of Experts model with 552 billion parameters. It features an asymmetric causal encoder-decoder architecture, with only 8 billion input activation parameters and 16 billion for output, along with native multimodal capabilities. By significantly compressing the KV cache, it reduces costs for agent scenarios. The API is now live, and Tencent's WorkBuddy, CodeBuddy, and OpenCode have fully integrated the model.
Source: DeepSeek
Read on 88z.aiQoder International adds the built-in Sonus model, focusing on long autonomous tasks and computer operations. Combined with desktop version 0.2.3 optimizations, the model can directly control computers and professional software to automatically execute complex jobs like programming and financial modeling.
Source: Qoder
Read on 88z.aiHiDream's content creation agent vivago R1 has launched globally, with its domestic version 'Gou Da' simultaneously upgraded. As the world's first infinite-duration creation agent, it supports one-time output of 5-minute high-quality videos with continuous extension capabilities, achieving an 85% effective usability rate. The product adopts a chat-first mode, driven by both foundation models and agents to automatically decompose scripts, storyboards, and sound effects.
Source: HiDream智象未来
Read on 88z.aiOpenAI has released ChatGPT Images 2.5, significantly improving detail, editing precision, and speed, with a 50% reduction in latency. New features like sketches and templates are now available, along with two new API models, Flare and Sunburst.
Source: OpenAI
Read on 88z.aiXiaomi has launched invitational testing for MiMo Desktop, entering the desktop Agent market to compete with Tencent, Alibaba, and ByteDance. Users can provide feedback via real tasks and access MiMo-X-Pro-Preview and MiMo-X-Flash-Preview for free on a limited basis.
Source: Xiaomi MiMo
Read on 88z.aiAnt Bailling releases its first native multimodal large model, Ling-3.0-flash-VL, with 124B total parameters and 5.5B activated per inference. It supports image, text, video, and 256K context, emphasizing low latency and cost.
Source: 百灵大模型
Read on 88z.aiMeta has launched Muse, its first personal AI agent capable of autonomously performing tasks such as sending emails, selling cars, and booking travel on behalf of users. The product is now available in the United States with a subscription fee. Muse can access sensitive data including emails, calendars, payments, and health records. Users can choose which apps to connect and retain the right to revoke access at any time.
Source: IT之家
Read on 88z.aiAnt Group's InclusionAI has open-sourced LLaDA-Image, a 6B-parameter text-to-image model. Featuring self-conditioned learning, unified generation and editing, and TwinFlow distillation that reduces steps from 50 to 2–4, it ranks first among open-source models in both Chinese and English on Qwen-Image-Bench, with a total score between GPT Image 1 and Imagen 4.0 Ultra.
Source: 蚂蚁开源
Read on 88z.aiDeepSeek has announced the internal testing of the intermediate version of DeepSeek V4.1 Flash. Featuring a new architecture with native multimodal support, the model offers enhanced capabilities, faster speed, and lower costs. Developers can access it by updating the model name while keeping the original base URL. Billing matches DeepSeek V4 Flash, with a limit of 20 concurrent requests per account. The model will automatically expire on September 10.
Source: 智东西
Read on 88z.aiFrench AI startup Mistral AI announced the completion of a €3 billion Series D funding round, reaching a post-money valuation of over €21 billion. Led by Samsung Electronics and others, the funds will support advanced research.
Source: IT之家
Read on 88z.aiModelBest and OpenBMB have released MiniCPM5-2B, an edge base model that ranks first among open-source models under 4B parameters on the AA leaderboard. It achieved a comprehensive score of 23 and an Agentic Index of 20. The project also open-sourced UltraData, Meshy framework, and JustRL II, with Day0 adaptation for Intel, Rockchip, and Arm.
Source: 面壁智能
Read on 88z.aiTencent Hunyuan announces an upgrade to the Hy4 preview model, focusing on optimizing long reasoning and excessive self-verification in complex tasks. The new version is fully released after dual monitoring via benchmark metrics and human evaluation. It significantly reduces task rounds and token consumption while maintaining task performance, thereby improving overall efficiency.
Source: 腾讯混元
Read on 88z.aiMusk's xAI has expanded the Grok Bot platform to iPad and Android systems, achieving coverage across Mac, iPhone, iPad, and Android devices. The platform supports 24/7 collaborative teamwork among AI agents, featuring cross-application login, context retention, and information sharing among multiple bots.
Source: IT之家
Read on 88z.aiMetaso has launched the 'Zhihui Yueyin' WeChat mini-program, an AI mental health app that uses structured card-based Q&A interactions combined with professional psychological logic to provide personalized emotional guidance and cognitive suggestions.
Source: 智慧月隐
Read on 88z.aiiFlytek has released the Spark X2.5 large model, trained entirely on domestic computing power using a MoE architecture with 293B-A30B parameters. The new version emphasizes enhanced code generation and agent capabilities, while also improving general performance in mathematics and understanding, continuing to overcome technical challenges in long-sequence training and inference on domestic hardware.
Source: 科大讯飞
Read on 88z.aiAlibaba launches QoderWake 1.0, enabling end-to-end management of AI employees for human-AI collaboration and scalable digital capacity.
Source: Qoder
Read on 88z.aiNVIDIA announced the acquisition of AI startup Hugging Face for approximately $13 billion to accelerate the adoption of open-weight models and strengthen its market position. The deal includes $11.9 billion to investors and $1 billion for employee incentives, with Jensen Huang promising to maintain its open platform status without forcing NVIDIA compute usage.
Source: 财联社
Read on 88z.aiOpenAI has released GPT-6 Astra, its most powerful and widely deployed model to date. It is the first to meet critical-level cybersecurity standards, capable of autonomously discovering and exploiting unknown vulnerabilities. While security and anti-interference capabilities are enhanced, its ability to evade monitoring has increased, resulting in reduced observability.
Source: IT之家
Read on 88z.aiAlibaba and Taotian Group jointly released the E-Commerce Bench, an evaluation benchmark for e-commerce operations. It requires large models to simulate running an online store for one year using 100,000 yuan in capital based on real market data, covering the entire process including product selection, pricing, negotiation, inventory, and promotions. Tests showed that GPT-5.6 Sol led with a 14-fold asset increase. Models varied significantly in anti-fraud and cash flow management, with only Qwen
Source: 千问大模型
Read on 88z.aiMicrosoft has released its latest AI speech-to-text model, MAI-Transcribe-2. In FLEURS tests across 60 languages, it achieved an average word error rate of just 5.2%, significantly outperforming competitors like Gemini, GPT-Transcribe, and Whisper in both accuracy and speed. The model supports intelligent speaker differentiation, word-level timestamping, flexible switching between verbatim and clean styles, and natural code-switching in conversations.
Source: IT之家
Read on 88z.aiGoogle has released Gemini 3.8, featuring a general Flash version and a cybersecurity-focused Cyber version. The former enhances reasoning and engineering capabilities while maintaining previous speed and cost levels, while the latter focuses on vulnerability remediation and is accessible to certified personnel via the Fairwind Program.
Source: 机器之心
Read on 88z.aiMeta has released its new large language model, Muse Spark 1.3. Its coding capabilities surpass GPT-5.6 Sol and match Claude Fable 5.1. Optimized for long-cycle Agent tasks, the model can proactively correct plans and ask clarifying questions. API pricing remains at previous levels, with tool calls reduced by approximately 20% and token consumption lowered by 25%.
Source: IT之家
Read on 88z.aiAnthropic has released two new models positioned as the most advanced tools for coding and knowledge work. Claude Fable 5.1 is available to the general public, while Claude Mythos 5.1 is restricted to vetted organizations. Fable 5.1 outperforms its predecessor and GPT-5.6 Sol in multiple benchmarks.
Source: IT之家
Read on 88z.aiThe Kimi Open Platform announced an API update, natively supporting integration with Codex and Claude Code. The Kimi API is now compatible with both the OpenAI Responses API and Anthropic Messages API formats, allowing users to connect directly to models like kimi-k3 and kimi-k2.7-code without format conversion or local proxies.
Source: Kimi开放平台
Read on 88z.aiAlibaba releases the Qwen3.8-Max-0902 model, enhancing programming and office scenarios. The model tops the CodeArena frontend leaderboard with a score of 1691, featuring superior multi-step reasoning and app generation capabilities, at a cost of only $5 per million tokens.
Source: 千问AI平台
Read on 88z.aiWorld Labs, founded by Li Fei-Fei, has unveiled Atlas, the world's first multimodal world model. Built on a multimodal autoregressive diffusion Transformer architecture, it features pixel-level camera control, 3D reconstruction, and spatiotemporal simulation. Outperforming current state-of-the-art models, Atlas is now available for early access to select partners.
Source: 量子位
Read on 88z.aiMeta has released Muse Voice Transcribe, its first real-time audio perception model, integrating streaming speech recognition, speaker diarization, and endpoint detection. Supporting over 70 languages and multi-track transcription for more than 20 speakers, the model uses an adaptive latency mechanism to balance speed and accuracy, ranking first on the Artificial Analysis streaming speech-to-text leaderboard.
Source: IT之家
Read on 88z.aiAlibaba Cloud has launched the public beta of 'Wanwu Wujie,' an enterprise-grade agent collaboration platform supporting the Qwen3.8-Max flagship model. The platform breaks down complex tasks among agents with different roles, facilitating work through project spaces, group chats, and asset libraries. It features two main agents, 'Xiao Wan' assistant and 'Xiao You' project manager, with capabilities for organizational memory ownership, permission control, and heterogeneous agent integration, ser
Source: 阿里云
Read on 88z.aiTencent Hunyuan has released the Hy4 preview lite version, utilizing its self-developed Sherry sparse ternary quantization algorithm and MIX-STQ1_0 mixed-precision layer-wise quantization strategy to compress model weights from 1.5 TB to approximately 214 GB. This quantized version performs close to the original BF16 model in tasks such as long-text understanding, multi-turn context retrieval, and coding assistance.
Source: 腾讯混元
Read on 88z.aiThe first multimodal model in the DeepSeek V4 series, DeepSeek-V4-Flash-Vision-Exp, has been open-sourced under the MIT license. It natively supports image inputs in JPEG, PNG, GIF, and WebP formats, offering capabilities such as image description, text recognition, and chart analysis. The model shows significant improvement in visual Agent benchmarks, with multimodal Agent capabilities approaching Opus-4.8, while pure text task performance matches the official V4-Flash release.
Source: IT之家
Read on 88z.aiAI 3D company VAST has recently completed its Series B and B+ funding rounds, raising a combined total of approximately 3 billion yuan. The round was led by Matrix Partners China, with participation from industrial investors such as Perfect World, BlueFocus, and 37 Interactive Entertainment, as well as financial investors including CDH Investments, CICC, and CMC Capital. Existing shareholders also made oversubscribed follow-on investments. This brings VAST's cumulative financing to about 5 billi
Source: 智能涌现
Read on 88z.aiTencent's WeChat Vision Team has open-sourced WeMM-Embedding, available in 2B, 4B, and 9B sizes. It supports unified multimodal encoding and ranks first on the MMEB-v2 benchmark.
Source: WeMM-Embedding
Read on 88z.aiXiaohongshu's FireRed team has released FireRedAudio, a general audio language model based on the Qwen3.5 architecture. It supports six major tasks including ASR and audio QA, achieving leading results in multilingual evaluations.
Source: FireRedAudio
Read on 88z.aiLoopit has open-sourced Zing-0.5, a real-time interactive video world model based on Wan2.2-TI2V-5B with 5 billion parameters. It supports combined keyboard and text control, generates video at over 24 FPS on a single RTX 5090, and ranks first in the WBench benchmark for real-time world models.
Source: Zing-0.5
Read on 88z.aiAnt Group has released Ling-3.0-flash-Fin, its first finance-enhanced model. Built on a 124B total parameter and 5.1B activated parameter architecture, it focuses on improving retrieval, reasoning, modeling, and report writing in financial workflows, demonstrating strong performance in various financial benchmarks.
Source: 百灵大模型
Read on 88z.aia16z has announced the launch of the $1.1 billion Machine Age Fund, focusing on AI physical infrastructure including chips, memory, networking, data centers, robotics, and home AI devices. The fund notes that as AI shifts from chat to reasoning and coding, compute demand is growing exponentially, challenging existing supply chains and physical limits, necessitating restructuring.
Source: a16z
Read on 88z.aiOpenAI has announced the termination of its partnership with Cursor following its acquisition by SpaceX, with direct model access ending on November 12. Citing compliance concerns and the security needs of the upcoming Astra model, OpenAI is tightening distribution controls.
Source: 机器之心
Read on 88z.aiTencent Hunyuan has released its new flagship model, Hy4 preview, featuring 770B total parameters and 49B activated parameters with a 1M context window. The model excels in coding, office tasks, and scientific research, boosting inference throughput by 31.8% and achieving breakthroughs in molecular dynamics simulation speed and geometric problem-solving.
Source: 腾讯混元
Read on 88z.aiGoogle has released the Gemini Omni 1.1 Flash video generation model, supporting output resolutions up to 4K. The model allows users to extend existing videos in 10-second increments, with a maximum single video length of 40 seconds, and enables smooth transitions by specifying start and end frames. It offers a 60% increase in generation speed at one-third the cost, supports multi-tier output from 360p previews to 4K ultra-high definition, and can incorporate reference videos up to 3 seconds lon
Source: IT之家
Read on 88z.aiZhipu has released GLM-5.3-Flash, the first native multimodal model in the GLM-5 series. It matches the intelligence level of Claude Opus 4.8 at a fraction of the cost. The model uses a hybrid architecture to significantly reduce long-context processing overhead and natively supports vision, code, and GUI collaboration.
Source: 智谱
Read on 88z.aiAlibaba Tongyi has released Qwen3.8-Flash, a multimodal MoE model with 125B total parameters and 6B activated per token, supporting ultra-long contexts. With four major technical upgrades, training costs are significantly reduced, coding and office capabilities are enhanced, and competitive API pricing is offered.
Source: 千问大模型
Read on 88z.aiTencent Hunyuan has released the on-device translation model Hy-MT2-1.8B, which uses proprietary compression technology to drastically reduce model size while maintaining high-quality translation that outperforms several commercial APIs. The model has been optimized for x86 operators and successfully deployed for real-time danmaku translation on Bilibili live streams, supporting 33 languages with a processing time of only 500-800 milliseconds per item.
Source: 腾讯混元
Read on 88z.aiQwenWork International Edition has launched its public beta, offering web and PC access to overseas individual and enterprise users. Positioned as an all-in-one AI productivity platform, it integrates with tools like Slack and Notion, supports multiple languages and account registration methods, and plans to deepen connections with enterprise databases and workflows.
Source: 千问办公QwenWork
Read on 88z.aiGoogle releases Gemini 3.5 Transcribe, which automatically removes filler words and self-corrections. It is about 70% faster than Chirp 3, with a real-time error rate of 5.5%. The model supports 85 languages and multi-speaker audio, and is already used in Pixel 11 Gboard.
Source: IT之家
Read on 88z.aiPixVerse has released the technical report for PixVerse R2, building on R1 to further explore scaling paths for real-time world models. Using the Omni Causal AR unified architecture, it supports multi-modal inputs including text, reference images, audio, and motion, enabling continuous world operation with simultaneous generation, interaction, and evolution.
Source: 爱诗科技 AIsphere
Read on 88z.aiStability AI has announced the completion of a $76 million Series B funding round, with proceeds allocated to creative product development and research. Investors include EA, Sony Music, Universal Music, Warner Music, and AMD's venture capital arm.
Source: IT之家
Read on 88z.aiJimeng AI has launched the 'Jimeng Studio' brand to solicit film and TV projects, offering full-chain support. It also initiated a creator growth plan with funding and resources to aid high-quality AI film IP development.
Source: 即梦AI
Read on 88z.aiGoogle Labs has released Play with Putty, a real-time collaborative tool focused on the Vibe Coding experience. Users without programming skills can describe requirements in natural language, allowing AI to generate code and render it into functional tools or websites in real time. The tool supports multi-user collaboration via shared links with instant synchronization of changes, aiming to transform software development from individual coding into collective dialogue for rapid prototyping.
Source: Putty
Read on 88z.aiByteDance has launched 'Doubao Work', an AI office assistant for productivity scenarios, featuring full-stack content generation, complex workflow handling, and deep integration with the Feishu ecosystem.
Source: 豆包
Read on 88z.aiXiaohongshu's FireRed team releases the open-source model FireRedTTS3, based on RedAE and LLM-DiT architectures, supporting zero-shot cloning and precise editing across multiple languages with top benchmark performance.
Source: FireRedTTS3
Read on 88z.aiLiblibAI has launched Newtake, an overseas AI video creation platform focused on cinematic content. Leveraging an infinite canvas and node-based workflow, it integrates professional features like a director's console and character consistency management, while supporting the Seedance 2.5 model to lower barriers for high-end video production.
Source: Newtake
Read on 88z.aiChina Telecom Tianyi AI has launched the desktop version of its Xingchen Super Agent, TeleAgent, positioned as an AI office companion. It supports Windows 10 and MacOS 13 or later. The tool automates tasks such as file organization, Excel summarization, data analysis, PPT creation, and information gathering, allowing users to execute commands via voice or text.
Source: 中国电信天翼AI
Read on 88z.aiAlibaba Cloud announces the official launch of video generation model Wan3.0, supporting single-generation of 30-second videos and input from various document formats, with applications deployed across multiple industries.
Source: 阿里云
Read on 88z.aiThe Wan3.0 model has officially launched on the RunningHub platform. This model supports single-generation videos up to 30 seconds long and accepts inputs including text, images, audio, video, and various document formats, enabling one-click conversion of PPTs into videos. It demonstrates significant improvements in character and scene consistency, effectively addressing the 'AI face' issue. Currently, RunningHub's ecosystem products, such as rhTV, RHSTORY, and ComfyUI, have been fully adapted t
Source: RunningHub
Read on 88z.aiAlibaba Tongyi has released Qwen-UI-Agent, a foundation model for GUI agents designed for mobile, desktop, and web scenarios. The model outperforms leading flagship models in various benchmarks and features strict safety confirmation mechanisms to ensure secure user operations.
Source: 千问大模型
Read on 88z.aiOpenAI has officially open-sourced Codex Harness, its underlying execution framework for AI agents. The framework covers the complete execution loop, including task understanding, memory retention, tool invocation, and human approval. It supports integration into business systems via CLI, SDK, and app-server. Key features include sandboxed secure execution, native Human-in-the-Loop mechanisms, streaming event transmission, and a model-agnostic architecture, aiming to provide developers with a fl
Source: Harness
Read on 88z.aiSenseTime has officially open-sourced the lightweight native unified multimodal large model, SenseNova U1.5 Lite. The model focuses on real-world visual creation workflows, with enhanced capabilities in long instruction following, visual quality, text and layout control, native 4K high-resolution output, and precise visual editing.
Source: 商汤科技SenseTime
Read on 88z.aiAnt Group's Bailin large model open-sources Ling-3.0-tiny-base and Ling-3.0-flash-base, releasing weights for six key training stages. The Tiny version focuses on code and low-cost research, while the Flash version balances code, reasoning, and long-context capabilities.
Source: 百灵大模型
Read on 88z.aiLuming Robotics unveiled the Lumos NexCore embodied AI evolution engine at the 2026 World Robot Conference. Positioned as industrial 'skill infrastructure,' it integrates data, training, evaluation, encapsulation, and operations to enable cloud-like invocation and iteration of robot capabilities.
Source: 机器之心
Read on 88z.aiDeepSeek Harness releases version v0.1.0-rc.8, focusing on enhanced multimodal support with native image requests and mixed text-image inputs, while allowing Claude Code and Codex to be called as sub-agents. The update also improves tool concurrency and Windows terminal experience, fixes stability issues, and reinforces its role as a unified agent scheduling layer.
Source: 机器之心
Read on 88z.aiMiniMax has launched MiniMax Design, an AI Agent platform built on the H3 multimodal video model, enabling full-process automation from intent expression to final delivery.
Source: MiniMax 稀宇科技
Read on 88z.aiThe Zhipu GLM-5.3 API officially launched today, specializing in complex coding, cybersecurity, and long-horizon tasks. Scoring 60 on the AA Comprehensive Intelligence Index, it rivals closed-source flagships like Claude Fable 5 and GPT-5.6 Sol, while tying with Kimi K3 for the top spot among open-source models. GLM-5.3 achieves frontier-level performance with smaller parameter sizes and the lowest single-task costs, keeping pricing consistent with the previous generation. Model weights are sche
Source: 智谱
Read on 88z.aiMicrosoft has released the open-source multimodal model family Mage, covering core capabilities in visual understanding and image generation. Mage-VL significantly reduces visual tokens via Codec-Native encoding, boosting video inference speed and enabling real-time streaming processing. Mage-Flow focuses on high-resolution image generation and editing, achieving rapid output on a single A100 GPU through efficient sampling in its Turbo version, while supporting diverse resolution requirements.
Source: Mage
Read on 88z.aiCoze Desktop officially launches, enabling Agents to read/write local files, execute commands, and manage projects, with a new cloud drive for cross-device collaboration and mobile remote control.
Source: 扣子Coze
Read on 88z.aiCursor has launched Origin, an AI-native code hosting platform, opening its beta to paid users. The platform offers repository management, PR review, and two-way GitHub sync, while integrating third-party apps like Vercel and enabling built-in AI agents to modify code directly.
Source: Cursor
Read on 88z.aiHiDream Future has released HiDream-O1-World, a native fully multimodal interactive world model supporting text, image, and interactive inputs. It features roaming, editing, and interaction capabilities. Built on the proprietary UiT architecture, it achieves breakthroughs in spatiotemporal and physical consistency, topping the WBench Navi leaderboard on its debut. Applications include AI interactive movies and games, embodied AI simulation, and 3D scene production.
Source: HiDream智象未来
Read on 88z.aiKunlun Wanwei releases Mureka V9.5, built on the MusiCoT framework, with significant improvements in vocal realism and intent accuracy. It breaks through in Chinese traditional style and, iterated on over 25,000 user feedbacks, aims to create AI music worth looping.
Source: 昆仑万维集团
Read on 88z.aiAlibaba Qwen Office has open-sourced MyContext, a context infrastructure project that runs locally, automatically consolidating multi-source work data into traceable exclusive work profiles to help users build dedicated Agent assistants.
Source: 阿里云
Read on 88z.aiAI video generation unicorn Higgsfield has completed a $400 million funding round, reaching a post-money valuation of $5.4 billion. Investors include DST Global, Goldman Sachs, Liberty Global, and Intel. The company's annualized revenue has reached $700 million, growing approximately 34-fold in one year, with its valuation rising 315% in eight months. The platform currently serves over 30 million users across 238 countries and regions.
Source: 智东西
Read on 88z.aiAlibaba has launched HappyShrimp, an AI music model capable of end-to-end full song generation, including lyric writing, composition, arrangement, and vocals. By deeply understanding natural language, it allows users to create music by simply describing emotions or stories, lowering barriers and addressing issues like robotic vocals and mismatched lyrics.
Source: 阿里云
Read on 88z.aiAlibaba's Qwen-Audio-3.0 series voice models have launched on the Qwen AI Platform, featuring ASR, TTS, and real-time interaction models, with API support and deployment in multiple products.
Source: 千问AI平台
Read on 88z.aiSpaceX has finalized the $60 billion acquisition of Cursor, setting a record for the largest startup buyout in history. The team will join SpaceXAI to co-develop the Grok series and Cursor products. Cursor will leverage SpaceX's global largest GPU cluster for enhanced computing power, while xAI has launched Grok Bot, capable of autonomously handling multi-step tasks around the clock.
Source: IT之家
Read on 88z.aiGoogle has released Gemini 3.7 Flash, a model designed for coding and agent scenarios that significantly improves software engineering and web development capabilities. It is available at half price until the end of 2026 and has been integrated into Workspace tools via Gemini Spark.
Source: IT之家
Read on 88z.aiZhipu has released GLM-5.3, significantly enhancing intelligence and coding capabilities through improved post-training. It ranks first among open-source models in multiple benchmarks. The model also demonstrates strong cybersecurity performance, prompting Zhipu to launch the 'Open Source Shield' initiative for vulnerability detection.
Source: 智谱
Read on 88z.aiMiniMax has released the open-source music model MiniMax-Music3, capable of generating lyricized songs up to five minutes long. Using a hierarchical architecture with an 8B global model and a 0.6B local model, it outputs 32kHz stereo WAV audio while maintaining consistency in theme, rhythm, and vocal identity across full song structures.
Source: IT之家
Read on 88z.aiXiaohongshu's dots model lab has released dots3-note preview, the first open-source version of the dots3 series. With 280B total parameters and 16B activated parameters, it supports a 512K context window and offers multimodal understanding of text, vision, and speech, performing competitively in various benchmarks.
Source: 机器之心
Read on 88z.aiDeepSeek has officially released DeepSeek V4 Pro, available on its app, web interface, and API. The new version significantly enhances agent capabilities, showing strong performance in benchmarks like Terminal Bench and DeepSWE. The API now natively supports the OpenAI Responses API format for Codex compatibility, and the thinking mode offers three selectable levels: low, high, and max.
Source: DeepSeek
Read on 88z.aiThe DeepSeek Harness Developer Preview (v0.1) is now open for global testing and released as open source under the MIT license. Built on a 'everything is a plugin' architecture, it allows flexible combination of models, tools, and skills, offering four operation modes and full tool invocation capabilities.
Source: DeepSeek Harness 团队
Read on 88z.aiSpaceXAI has released Grok Bot, an AI agent system capable of independent cloud operation and multi-agent collaboration.
Source: APPSO
Read on 88z.aiSpaceXAI has officially released the Grok 4.6 model, significantly enhancing capabilities in long-duration agent tasks, complex interactions, and visual processing. It supports data research and codebase analysis. In the AA Intelligence Index benchmark, its performance matches that of GPT-5.6 Sol.
Source: IT之家
Read on 88z.aiXiaohongshu's dots team and Shanghai Jiao Tong University's X-LANCE Lab jointly open-sourced dots.tts, a 2-billion-parameter, fully continuous latent space autoregressive TTS base model. The model does not rely on discrete acoustic tokens, achieves SOTA on the zero-shot voice cloning benchmark Seed-TTS-Eval, and supports timbre fine-tuning and speech editing.
Source: 机器之心
Read on 88z.aiMiaowu Meoo has launched a memory feature that automatically organizes and saves long-term user information such as personal background and work habits for intelligent retrieval in future conversations. This feature includes two categories: personalized memory visible only to the user and application memory shared within teams.
Source: 秒悟Meoo
Read on 88z.aiOpen-world model company LTX has released LTX-2.5 with native ComfyUI integration. On dual NVIDIA GB200 systems, it generates a 10-second 720P video in just 6.8 seconds, while the API version outputs 1080P video in 23.7 seconds. The new model offers improved pixel fidelity, multi-shot consistency, and diffusion rendering quality, accessible via Hugging Face, ComfyUI, and API.
Source: IT之家
Read on 88z.aiThe Qwen team has open-sourced the Qwen3.8-2.4T-A95B model, marking the first time weights for a Qwen-Max level model have been released. This MoE model features 2.4 trillion total parameters with 95 billion activated per token and natively supports a 256K context window. It focuses on enhancing capabilities in coding, office tasks, scientific research, and long-cycle agents, leading in benchmarks like PaperBench and TerminalBench. The model supports deployment via frameworks such as SGLang and
Source: 魔搭ModelScope社区
Read on 88z.aiManus has announced its separation from Meta to resume independent operations. Due to cross-border regulatory reviews, both parties adjusted their structure. Manus requested some users to back up data by August 23 and will restore services on August 25, offering free access during the transition and return incentives.
Source: APPSO
Read on 88z.aiBilibili has open-sourced IndexTTS-2.5, an industrial-grade zero-shot voice cloning model. With only 0.8 billion parameters, it supports cross-language transfer across Chinese, English, Japanese, Spanish, and Arabic. The restructured inference architecture boosts speed by 2.28 times, achieving a real-time factor as low as 0.20. It also offers seamless speed adjustment from 0.5x to 2.0x and provides pronunciation intervention at three granularities: Pinyin, CMU phonemes, and Japanese Kana.
Source: IndexTTS-2.5
Read on 88z.aiDoubao has launched a special offer for college students, providing 2.5 times the free quota and discounted rates for the professional version to assist with complex tasks like research and creation. Users can verify their status via specific commands on the PC client or by linking their Douyin accounts.
Source: 豆包
Read on 88z.aiMicrosoft has released the upgraded programming model MAI-Code-1.1-Flash, featuring significantly enhanced coding capabilities and new native visual functions. The model shows a 22% improvement in Terminal-Bench 2.1 tests and a 15% boost in .NET tasks. Additionally, token generation speed is increased by 25% with a 25% reduction in consumption, while the price is cut to one-quarter of the original. The new model is gradually being integrated into Copilot platforms like VS Code and JetBrains, wit
Source: IT之家
Read on 88z.aiOpenAI has expanded its Daybreak cybersecurity defense service by adding Blue and Red tiers. The Red tier includes the new GPT-5.6-Cyber model, trained for security testing, which is currently available only to trusted partners such as Accenture and IBM.
Source: IT之家
Read on 88z.aiMeta has released Muse Glimmer, an open-source large language model with 30 billion parameters. Optimized for local deployment, it uses 4-bit quantization and DFlash speculative decoding to run smoothly on devices with 24GB VRAM, such as Macs or consumer GPUs, supporting offline multimodal perception, tool use, and code generation.
Source: APPSO
Read on 88z.aiZCode, under Zhipu AI, has undergone a comprehensive upgrade, introducing four core features: Goal, Subagents, Remote Control, and Off-Peak Tasks. The Goal mode enables autonomous breakdown and delivery of complex tasks, Subagents facilitate parallel multi-agent collaboration, Remote Control allows management via mobile or WeChat, and Off-Peak Tasks run without points during low-traffic periods. Meanwhile, GLM Coding Plan quotas have been uniformly reset, with all user balances fully restored.
Source: Zcode
Read on 88z.aiTencent has launched Xiao'e Desktop, a Windows customization tool offering free, ad-free access to millions of 4K wallpapers. It features a lightweight AI widget for tasks and file search, targeting students and individual users to complement its enterprise-focused WorkBuddy.
Source: 小鹅桌面
Read on 88z.aiAlibaba has released CosyVoice Studio, the first one-stop AI voice platform in China. Built on the self-developed Qwen-Audio model, it integrates three core modules for voice recording, agents, and audio creation.
Source: 阿里云
Read on 88z.aiNVIDIA has released Alpamayo 2 Super, an open-source autonomous driving model built on Cosmos 3 Super Reasoner. Licensed under OpenMDW 1.1 for commercial use and fine-tuning, it ranks top in benchmarks like LingoQA. The model outputs five result types, including trajectory planning and causal reasoning, ensuring transparent and verifiable decision-making.
Source: Super
Read on 88z.aiCloudflare has released Kitesurf, a cloud-based browser designed specifically for AI agents. Built on the Workers platform, it aims to reduce computational costs while optimizing performance and security.
Source: IT之家
Read on 88z.aiTencent Hunyuan's 3D research team has released WorldClaw, an agent-based framework for generating open 3D worlds. It converts text prompts into explorable and editable 3D environments, maintaining terrain consistency while adding local details, and supports object-level editing and game engine integration.
Source: WorldClaw
Read on 88z.aiAlibaba has opened public testing for its new generation video generation model, Wan3.0. The model can generate 30-second videos in a single run, supporting continuous camera movement and single-take narratives. In addition to text, image, audio, and video inputs, it now supports document formats such as doc, xls, ppt, pdf, and md for the first time, allowing office materials to be directly converted into teaching courseware or product demonstration videos. Portrait generation aims for unique in
Source: 阿里云
Read on 88z.aiOpenAI announced that text chats with ChatGPT are no longer restricted for free users, with the default model upgraded to GPT-5.6 Luna, which outperforms GPT-5.5 Instant. Free and Go subscribers can also use a new thinking button for complex problems, but high-compute features like file uploads, image generation, and voice conversations remain limited. Paid users will receive the more reliable and concise GPT-5.6 Sol along with a five-level thinking intensity slider.
Source: 财联社
Read on 88z.aiAlibaba's new generation video generation model Wan3.0 has launched public beta on the Duiyou platform, supporting 30-second video generation and multimodal inputs, with significantly improved character consistency and editing capabilities.
Source: 堆友
Read on 88z.aiVolcano Engine has officially launched the Seedance 2.5 API, delivering significant improvements over version 2.0 in instruction following, long-form narrative, human-like realism, and audio-visual quality. The service natively supports direct generation of 30-second videos, allows up to 50 full-modality reference materials, and is compatible with over ten languages. The model is now simultaneously available on platforms including LibTV, Qixiang AI, and TapNow.
Source: 火山引擎
Read on 88z.aiAnt Bailin has released Ling-3.0-tiny, a native hybrid reasoning model with 7.9B total parameters and 1.3B activated per token. Optimized for mathematical reasoning and edge deployment, it supports tool calling and fully local offline inference.
Source: Ling-3.0-tiny
Read on 88z.aiRabbitPre Intelligence launches RabbitVis, a one-stop AI design tool based on its self-developed visual large model, deeply integrating generation with layered editing.
Source: 量子位
Read on 88z.aiMeta has released the beta version of Muse Code, its first coding AI agent tool, led by Chief AI Officer Wang Tao. Built on the Muse Spark 1.2 model, it aims to handle complete software engineering tasks within large codebases, directly challenging Anthropic's Claude Code and OpenAI's Codex.
Source: IT之家
Read on 88z.aiAlibaba Cloud has launched the One Key MCP service, enabling developers to call all ecosystem partners' MCP services with a single Alibaba Cloud Bailian API key. Compatible with mainstream Coding Agents such as Qoder, Codex, Claude Code, and Cursor, it significantly simplifies access, authentication, and billing management for multiple MCP services. Users can log in to the Alibaba Cloud Bailian platform and directly activate and call relevant services in the MCP Square using their existing API k
Source: 阿里云
Read on 88z.aiTencent has launched Omega, an AI-native data analytics product, with its beta version named 'Marmos'. Users can generate data dashboards via natural language, supporting interactive modifications and dynamic refreshes. The product utilizes QueryRegistry and DTBridge architectures, combined with multi-layer permission verification and AST read-only checks, aiming to resolve data stagnation and security issues in traditional AI dashboards.
Source: Omega
Read on 88z.aiAlibaba's Qwen team and XLang Lab have jointly released Qwen-CUA, a native computer use agent. Built on a mixture-of-experts architecture, it perceives interfaces via screenshots and directly outputs keyboard and mouse events to control software without relying on underlying metadata. It performs well on the OSWorld-Verified benchmark, with the trillion-parameter version scoring 87.6.
Source: Qwen-CUA
Read on 88z.aiTencent Hunyuan has released the new Hy ASR 3.0 preview speech recognition model. Built on the Hy3 large language model, it deeply integrates speech recognition with deep semantic understanding. The model achieves top-tier industry accuracy in Mandarin, English, and Cantonese scenarios, featuring intelligent error correction in complex contexts.
Source: 腾讯混元
Read on 88z.aiAlibaba's Qwen image generation model, Qwen-Image-3.0, has launched on the Qwen AI platform and is freely available to all users. The flagship and standard versions have simultaneously opened their APIs. In the latest text-to-image rankings on the international evaluation platform Arena.ai, the model ranks first in China and fifth globally.
Source: 阿里云
Read on 88z.aiByteDance Seed launches SeedRealtime, a native audio-video full-duplex large model that integrates audio, video, and text via a unified architecture for real-time multimodal interaction.
Source: 字节跳动Seed
Read on 88z.aiMistral AI has released Shieldstral, an open-source multimodal content moderation model supporting 12 languages and runnable on a single 16GB GPU. The model achieves state-of-the-art performance in multimodal moderation, rivaling competitors seven times its size. It innovatively inputs moderation policies as natural language questions, outputting calibrated safety scores via a yes/no Q&A mechanism, suitable for prompt classification, response moderation, and toxicity detection.
Source: IT之家
Read on 88z.aiJD.com has open-sourced its self-developed real-time streaming video editing model, JoyAI-Video-Edit. The model supports editing while playing, achieving 30 frames per second inference at 720P resolution and handling videos of any length. It outperforms similar streaming models like SANA Streaming and LiveEdit in OpenVE-Bench evaluations, with editing quality matching that of offline models.
Source: 京东黑板报
Read on 88z.aiTencent Hunyuan, together with Tsinghua AIR and Southeast University, has released E-Bench, an agent evaluation benchmark. Built on fully synthetic virtual environments modeled after Honor of Kings, QQ Music, and Tencent Meeting, it includes 323 state-change tasks and over 76,000 data entries. By incorporating information and tool gaps, it prevents models from relying on memory shortcuts. Evaluations show that 11 leading models achieved an average success rate of only 54.56%, with the top model
Source: 腾讯混元
Read on 88z.aiSenseTime has signed a strategic cooperation framework agreement with the Ministry of Education's Educational Technology and Resource Development Center. The partnership aims to promote the application of AI in public education services, deepen the integration of educational technology and teaching resources, and empower the intelligent construction of national-level education service platforms.
Source: 商汤科技SenseTime
Read on 88z.aiSenseTime has open-sourced SenseNova U1.5-Lite-Preview, a lightweight native unified multimodal model based on the NEO-Unify architecture, which integrates visual understanding, reasoning, generation, and editing using only 8B-MoT parameters.
Source: 商汤科技SenseTime
Read on 88z.aiModelBest and OpenBMB have open-sourced ForgeStencil, the world's first AI-driven automated Stencil optimization system. By leveraging the collaboration of two agents, Kernel Agent and App Agent, the system achieves a fully automated workflow from operator research to application deployment without human intervention. ForgeStencil demonstrates high efficiency by completing end-to-end optimization for over 100 real-world industrial and scientific software programs per week, covering eight major i
Source: 面壁智能
Read on 88z.aiPacini Perception Technology announced the completion of a 1 billion yuan strategic financing round, bringing its total funding to 3.5 billion yuan, setting a new record in the global haptic perception sector. The round was co-led by major consumer electronics and semiconductor giants, BOC International Investment, and Kunpeng Fund, with existing shareholders increasing their stakes. Focusing on full-stack self-developed embodied perception technology, the company aims to build physical interact
Source: 帕西尼感知科技
Read on 88z.aiAlibaba Cloud has launched 'Wanwu Wujie', a B2B AI office platform that uses digital employee teams to manage complex projects, while also supporting lightweight scenarios for freelancers.
Source: IT之家
Read on 88z.aiAlibaba Cloud releases Qwen3.8-Max with 2.4 trillion parameters and 95B activated parameters, supporting million-level context. The model is upgraded for coding, office, and research tasks, with weights open-sourced next week and domestic API input priced at 12 yuan per million tokens.
Source: 千问大模型
Read on 88z.aiMiniMax has officially open-sourced its new general video model, MiniMax H3, which supports multimodal input and generates high-quality videos. The model consists of three modules, with the base module already open-sourced and supporting local deployment on mainstream frameworks.
Source: MiniMax 稀宇科技
Read on 88z.aiGenspark founder Jing Kun unveiled GenOffice at the AGI Playground 2026 conference, positioning it as the world's first full-featured open-source AI Office suite. The client supports Windows and Mac, offering editing capabilities for documents, spreadsheets, slides, and PDFs, while being free, ad-free, and open-source.
Source: 极客公园
Read on 88z.aiDiffusion language model developer Westlake Xinchen has completed a Series B+ funding round worth hundreds of millions. Led by GF Xinde, the investment will support new model R&D, multimodal product iteration, and overseas expansion.
Source: 西湖心辰
Read on 88z.aiByteDance's Seed team launches Seedance 2.5, doubling single-generation duration to 30 seconds with extension support, while enhancing multimodal referencing and professional editing capabilities.
Source: 字节跳动Seed
Read on 88z.aiZhipu GLM Coding Plan subscription has reopened with a transparent points system and off-peak deductions. Existing users retain benefits and can renew at V2 pricing, while HighSpeed edition applications are now open.
Source: 智谱开放平台
Read on 88z.aiMiniMax has released MiniMax H3, a general all-modal generative model that supports unified understanding and generation of text, images, video, and audio. It can output 15-second 2K resolution native stereo audio-video content. The model breaks task and modality boundaries, excelling in instruction following, motion transfer, and multimodal editing, making it suitable for commercial scenarios such as advertising, e-commerce, and gaming.
Source: MiniMax 稀宇科技
Read on 88z.aiDeepSeek has launched the public beta for the official DeepSeek-V4-Flash API. This release features significantly enhanced agent capabilities, outperforms V4-Pro-Preview in benchmarks, natively supports the Responses API format, and is optimized for Codex, while maintaining the original model structure with only post-training updates.
Source: IT之家
Read on 88z.aiTencent WorkBuddy version 5.3.5 collaborates with Tencent Docs to introduce a human-AI co-writing feature, enabling users and AI to edit the same document in real time. The tool allows modifications within selected areas, supports various mainstream formats as well as local and cloud documents, and facilitates collaborative creation with multiple users and AI with real-time synchronization.
Source: 腾讯
Read on 88z.aiLibTV and MiniMax have jointly released the new multimodal video model MiniMax H3, which supports 5 to 15-second video generation, 24 FPS output, and native dual-channel audio. The model features unified understanding of text, images, audio, and video, enabling an integrated creation workflow from reference materials to final editing, targeting commercial scenarios such as short dramas, advertising, and film production.
Source: 哩布哩布AI
Read on 88z.aiWorld Labs, founded by Li Fei-Fei, has acquired SceniX, marking a shift in physical AI training from collecting real-world data to generating virtual environments. SceniX specializes in reconstructing real scenes into interactive physical digital twins, creating controlled variations based on geometry, materials, and layout to build infinitely generalizable virtual worlds from limited real data.
Source: 量子位
Read on 88z.aiMoonshot AI's Kimi has completed a new funding round, raising $3.5 billion and reaching a valuation of $35 billion, while seeking new financing with a pre-money valuation of approximately $50 billion. Kimi open-sourced the K3 model weights on Hugging Face, featuring 2.8 trillion total parameters and 104 billion activated parameters, making it the world's first open-source model at the 3-trillion-parameter level. Based on MoE architecture, K3 has been adapted by multiple platforms and is priced s
Source: 第一财经
Read on 88z.aiGoogle DeepMind has launched Lyria 3.5, a new generation AI music generation model on Flow Music, featuring core upgrades in musicality, lyric quality, vocal expressiveness, and creative control. The model generates more natural and complex melodies, significantly improves lyric adherence, delivers emotionally rich and clear vocals, and supports easy adjustment of rhythm and duration.
Read on 88z.aiByteDance has initiated a restructuring of its AI business, merging the Lark product team into Doubao under Zhao Qi's leadership. The Lark GTM team and Volcano Engine are combined into a Creativity Service Platform led by Tan Dai. Data shows Doubao has 382 million monthly active users, and ByteDance's large model ARR has reached $4 billion.
Source: 财联社
Read on 88z.aiTencent Hunyuan team open-sources speculative decoding framework AngelSpec, covering the full pipeline from drafter training and architecture design to online deployment. It also releases weights and training code for Hy3-A21B's MTP and DFly drafters. DFly achieves SOTA at 4-64 concurrency, accelerating 1.98-2.40x over autoregressive baselines on average, with a peak of 2.86x in code and math. D-cut boosts throughput by an additional 15.7% under high concurrency, while combining MTP with TTT rai
Source: 腾讯混元
Read on 88z.aiSpaceXAI has launched Build mode for Grok, allowing users to turn ideas into publishable products with custom domains via prompts. It supports generating websites and apps, enabling code-free adjustments and source code export.
Source: IT之家
Read on 88z.aiXinyan Group has launched AgentLink, a cross-agent terminal that integrates dispersed AI agents like Claude, Codex, and OpenClaw into mobile devices. Users can remotely control and collaborate on tasks by scanning codes. The tool displays agent thinking processes and progress in real time, allowing for human confirmation at key nodes. Its unique cloud note relay feature enables closed-loop collaboration among multiple agents. Compatible with dual platforms, AgentLink offers a new solution for un
Source: AgentLink
Read on 88z.ai360's Nano Team has launched Nano Work, a new enterprise agent platform covering six work scenarios. It uses multi-agent collaboration to deliver finished outputs like PPTs, features five technical highlights, and supports 24/7 cloud execution.
Source: 纳米Work
Read on 88z.aiHuawei teams jointly launched JiuwenSwarm, an open-source unified AI workbench for HarmonyOS PC, supporting multi-agent collaboration across scenarios and introducing the new HITS human-agent collaboration paradigm.
Source: 量子位
Read on 88z.aiSenseTime has released Seko 3.0, upgrading it from a creation tool to a creator growth platform. This version pioneers the 'AI Video Dream Factory,' linking infinite canvas assets and the Skill ecosystem via Agent intelligent workflows. It encapsulates top creators' experience into callable professional capabilities, enabling systematic film-grade content production.
Source: 商汤科技SenseTime
Read on 88z.aiMoonshot AI has released the Kimi K3 model weights, technical report, and three key infrastructure technologies. The model features a 2.8 trillion parameter MoE architecture with native visual understanding and a one-million token context window, offering significantly improved efficiency. Additionally, MoonEP, FlashKDA, and AgentEnv have been open-sourced, covering critical links from training communication to distributed reinforcement learning environments.
Source: 月之暗面Kimi
Read on 88z.aiRenmin University of China and Ant Group have jointly launched SearchOS, an open-source multi-agent search framework designed to provide system-level collaboration infrastructure for long-horizon complex information retrieval. By abstracting task progress, evidence relationships, and completion status into a system state shared across agents, the framework enables coordinated division of labor under unified scheduling, effectively preventing redundancy and memory loss.
Source: 机器之心
Read on 88z.aiMicrosoft has launched its first cybersecurity AI model, MAI-Cyber-1-Flash. It performed well in benchmarks, is integrated into the MDASH multi-agent security system, handles most security tasks efficiently, and significantly reduces costs.
Source: IT之家
Read on 88z.aiMidjourney releases the V8.2 model, significantly improving prompt understanding, style consistency, and text rendering. It optimizes complex scene analysis and Eastern aesthetic expression, greatly increasing the success rate of single-generation outputs.
Source: V8.2
Read on 88z.aiAnthropic has released Claude Opus 5, which matches the intelligence of Fable 5 at half the cost. It sets new records in programming and knowledge work benchmarks and is now the default model for Claude Max.
Source: 机器之心
Read on 88z.aiAndrew Ng has open-sourced OpenWorker, a desktop agent that focuses on delivering completed work by autonomously using local files and office tools. It prioritizes local processing and privacy, keeping data on user devices while allowing model switching. It currently supports macOS, with a Windows beta available.
Source: 量子位
Read on 88z.aiMeituan has launched CatPaw, a full-scenario AI Agent platform offering ready-to-use intelligent workbenches and enterprise-grade development hosting. Deeply integrated with local life industry insights, the platform supports multi-terminal 24/7 operation and has been widely adopted internally. It features core capabilities such as multi-Agent collaboration, plug-and-play skills, and long-term memory.
Source: MeituanCatPaw
Read on 88z.aiAlibaba has launched the beta version of 'Qwen Office', supporting Windows, macOS, and HarmonyOS, with six core features including AI writing and document generation.
Source: IT之家
Read on 88z.aiGlobal AI interactive entertainment platform Haiyi announced the completion of its Series B financing round, raising over 100 million yuan. The round was co-led by Visual China Group, Huagai Chuangying, and Vertex Ventures. Haiyi has over 65 million registered users, with more than 90% from overseas markets, and maintains a gross margin exceeding 40%. The platform has accumulated over 2 million AI-native creative assets, establishing a comprehensive AI IP ecosystem comprising SeaArt for characte
Source: Z Potentials
Read on 88z.aiBlack Forest Labs has released FLUX 3, a multimodal foundation model that jointly learns images, video, and audio within a unified architecture. Built on Self-Flow technology, it supports text-to-video and image-to-video generation, producing videos with native audio up to 20 seconds long in a single pass, while demonstrating strong performance in robotic manipulation tasks.
Read on 88z.aiTRAE Work has upgraded its built-in Seedream 5.0 and Seedance models, offering them for free. Users can generate images and videos directly within workflows using natural language, with content seamlessly embedded into subsequent tasks.
Source: TRAE.ai
Read on 88z.aiMicrosoft has officially released two self-developed AI models, MAI-Image-2.5-Pro and MAI-Voice-2-Flash, opening them for public preview. The former is the company's most accurate image generation model to date, supporting main visual creation, detail editing, and text rendering. The latter is optimized for high-concurrency voice scenarios and supports 15 languages.
Source: IT之家
Read on 88z.aiXiaohongshu's dots infra team has open-sourced BigMac, a multimodal large model training framework that balances computational efficiency and memory usage via nested pipeline design, achieving significant speedups and already deployed in production.
Source: 小红书技术REDtech
Read on 88z.aiAlibaba Cloud has released the open-source document parsing model OvisOCR2, which set a new record of 96.58 on the OmniDocBench v1.6 leaderboard, becoming the first end-to-end model to surpass traditional pipeline methods and take the top spot.
Source: 阿里云
Read on 88z.aiClaude Cowork has launched the Record a Skill feature, enabling users to automatically generate reusable AI skills by recording screen actions and voice explanations. This tool simplifies workflows and is available to Pro, Max, and Team subscribers via the plus menu in the desktop app's Cowork mode.
Source: APPSO
Read on 88z.aiOpenAI has officially launched OpenAI Presence, an enterprise-grade AI agent platform designed to help businesses efficiently deploy and manage AI agents for automating tasks such as customer support and sales. The platform deeply integrates with internal corporate data, management protocols, and existing software, while offering management tools including simulation testing, security protection, human review, and AI automated scoring, along with integrated voice and chat technologies.
Source: IT之家
Read on 88z.aiMicrosoft has released Mage-Flow, an open-source series of lightweight multimodal image generation models based on flow matching architecture with only 4 billion parameters. The series includes Base, RL, Turbo, and Edit versions, supporting text-to-image generation, local editing, and style transfer. It runs efficiently on consumer-grade GPUs, suitable for e-commerce, advertising, and game development.
Source: Mage-Flow
Read on 88z.aiTencent Cloud has launched CodeBuddy NPC, a cloud-based agent positioned as an 'AI employee living in the cloud.' Developers can assign tasks by mentioning the NPC in issues, allowing it to autonomously handle everything from planning to code delivery. It supports multi-agent collaboration for complex tasks and has reduced initial token consumption by over 90%.
Source: 腾讯云
Read on 88z.aiHiDream.ai has announced the completion of a 1.5 billion RMB Series C funding round, co-led by the Sichuan Revitalization Science and Technology Innovation Fund of the National Social Security Fund and ICBC Capital. With cumulative financing exceeding 2.1 billion RMB in the past three months, the company has officially entered unicorn status. The funds will accelerate its technological evolution from multimodal large models to native omni-modal world models.
Source: HiDream智象未来
Read on 88z.aiSenseTime released SenseNova U1 Pro at WAIC 2026, an AI image generation system designed for complex multimodal tasks. The model supports native 8K output and possesses interleaved text-image reasoning capabilities, allowing it to autonomously handle the entire workflow from sketching to adjustment. Tests show stable performance in scenarios such as ultra-long aspect ratios, complex posters, and movie posters, with clear text and complete composition.
Source: 量子位
Read on 88z.aiGoogle has released three new AI models focusing on performance, cybersecurity, and agent applications. Gemini 3.6 Flash shows improved results in coding and finance benchmarks at a lower cost, while the 3.5 Flash Cyber version targets vulnerability remediation and is restricted to government entities and trusted partners to prevent misuse.
Source: IT之家
Read on 88z.aiTencent's AI creative agent platform, Miora, has officially launched at full scale. Positioned as an AI creative studio with memory, requirement understanding, and multi-agent collaboration capabilities, it allows users to state their goals. The system then automatically breaks down tasks and coordinates experts to handle various materials while maintaining consistent style.
Source: 腾讯
Read on 88z.aiSenseTime has launched the 'Kapi Family' AI application matrix, centered on personal intelligence. With over 40 million cumulative users and annual growth exceeding 500%, the suite includes apps for health advice, AI photography, and financial management.
Source: 商汤科技SenseTime
Read on 88z.aiModelBest released the open-source embodied AI model series MiniCPM-Robot at WAIC 2026, featuring the general manipulation model RobotManip and the tracking navigation model RobotTrack. The former achieves high performance with small parameters and supports native memory, while the latter is the first tracking model supporting local offline deployment.
Source: 面壁智能
Read on 88z.aiKunlun Wanwei releases the open-source interactive world model Matrix-Game 3.5, featuring a lightweight architecture and automated data pipeline for real-time interaction on a single GPU.
Source: 昆仑万维集团
Read on 88z.aiTencent Hunyuan has released Hyra, a general scientific discovery agent supporting recursive self-improvement. The model excels in AI R&D benchmarks and math problems, setting new records while achieving breakthroughs in hardware design and drug discovery.
Source: 腾讯混元
Read on 88z.aiYuanshi Technology has released Wenxiaobai 5 Pro, positioned as a new generation long-content generation engine focusing on length, stability, and traceability. The product addresses common issues in AI writing such as data hallucinations, inconsistent settings, and loose structures by employing database citation tracking, project memory mechanisms, and long-text structure preservation technologies to ensure logical coherence and framework integrity for documents spanning tens of thousands of wo
Source: 问小白
Read on 88z.aiAlibaba has officially launched Qwen-Audio-3.0-TTS, a large speech synthesis model featuring Flash for real-time interaction and Plus for high-quality generation. The model delivers systematic improvements in fine-grained tag control, natural language style definition, multilingual and dialect coverage, and acoustic robustness. It supports precise tone and emotion regulation via structured tags in text, covering 16 languages and 20 dialects.
Source: 千问大模型
Read on 88z.aiMeituan's LongCat team released the LoHoSearch benchmark, generating difficult questions using Wikipedia knowledge graphs. Tests showed GPT-5.5 achieved only 34.74% accuracy, exposing bottlenecks in search agents for complex reasoning tasks.
Source: 龙猫LongCat
Read on 88z.aiAlibaba's Qwen has launched the Qwen3.8-Max preview, with the full version set for release and open-sourcing soon. The model features 2.4 trillion parameters and is described by officials as potentially the most powerful model aside from Fable 5. The preview is currently available on Alibaba Token Plan, Qoder, and QoderWork platforms.
Source: 千问大模型
Read on 88z.aiQuwan Technology has released Tempolor v4.7, integrating the conversational music agent Tunee and opening its API. The new version focuses on controllable editing, upgrading Remix and Cover capabilities with fine-grained audio adjustments. Built on a fourth-generation layered progressive architecture, the model offers a 20% performance boost and outputs 48kHz stereo audio.
Source: 趣丸科技
Read on 88z.aiByteDance Seed releases Seed Audio 1.0, unifying modeling for vocals, sound effects, and ambient sounds to support end-to-end film-grade audio creation and multilingual generation.
Source: 字节跳动Seed
Read on 88z.aiMeshy has announced the completion of a Series B financing round of nearly $400 million, achieving a post-money valuation exceeding 10 billion yuan, setting a new record for the highest valuation in the global AI 3D sector. Over the past year, its revenue has increased twelvefold, with over 12 million registered users and more than 100 million models generated cumulatively. Half of the world's top ten technology companies have become its clients.
Source: 智东西
Read on 88z.aiMoonshot AI releases Kimi K3, the world's first open-source model in the 3-trillion parameter class with 2.8 trillion parameters, supporting a million-token context window and native visual understanding, with leading performance in multiple benchmarks.
Source: 月之暗面 Kimi
Read on 88z.aiGoogle has announced that its AI note-taking assistant, NotebookLM, is being renamed Gemini Notebook to unify its AI product branding. The product will continue to operate independently while integrating deeply into the Gemini app and Google Search ecosystem. The new version supports secure cloud computers, featuring native code writing and execution capabilities for complex data analysis based on data sources.
Source: IT之家
Read on 88z.aiNubia's AI companion robot iMoochi is now officially available for 1699 yuan. First unveiled by ZTE at MWC26, it features a cute plush design and lifelike interactions. Using touch sensors and AI algorithms, it responds dynamically and develops a unique personality over time, aiming to provide emotional comfort to urban individuals living alone.
Source: IT之家
Read on 88z.aiNVIDIA has released the open-source Nemotron 3 Embed text embedding model series, optimized for retrieval-augmented generation and agent search. The series offers 8B and 1B parameter sizes, supporting a 32k context window and 34 languages. The flagship 8B model ranks first on the RTEB multilingual retrieval benchmark with a score of 78.5%, while the 1B version retains 95% retrieval accuracy while significantly reducing inference costs through pruning and distillation.
Source: Embed
Read on 88z.aiBaidu AI Cloud released Miaoda 3.5 at WAIC2026, featuring a built-in SEO Agent and global debut iOS packaging capabilities for no-code app generation directly to iPhone. The update also introduces shared backends and multi-environment isolation, creating a closed loop from creation to multi-end delivery to promote accessible no-code development.
Source: 百度智能云
Read on 88z.aiMultimodal AGI firm Prana Labs has closed a seed round of nearly ten million US dollars, with investment from Yuan Sheng Capital, XVC, Creekstone, and 37 Interactive Entertainment. The company focuses on multimodal foundation models, aiming to build an interactive world generation system with physical constraints to train next-generation general AI through trial and error in virtual environments.
Source: Z Potentials
Read on 88z.aiElon Musk's xAI has announced the open-sourcing of its AI coding agent tool, Grok Build, with source code released on GitHub. Integrating the Grok model, the tool covers the entire development workflow from planning and search to coding, testing, and Git commits. The release includes the agent pipeline, code tools, terminal UI, and extension system, supporting fully local-first execution and custom inference.
Source: IT之家
Read on 88z.aiOpenAI and peripheral manufacturer Work Louder have jointly launched Codex Micro, the first hardware product for Codex. Priced at $230 and now available for pre-order, the keyboard is designed specifically for Codex. It uses RGB lighting to map AI agent status in real time, features a knob to adjust inference intensity, and includes a joystick and physical keys for workflow distribution and high-frequency decisions, allowing users to track progress without switching windows.
Source: APPSO
Read on 88z.aiEdge AI unicorn ModelBest has completed a new funding round, with its valuation surpassing 20 billion yuan, making it the highest-valued unicorn in China's edge intelligence sector. Incubated from Tsinghua University's NLP Lab, its core MiniCPM series has released its latest open-source version and achieved commercial deployment across automotive, mobile, and PC sectors.
Source: 智东西
Read on 88z.aiAlibaba Tongyi Lab releases Qwen-Audio-3.0-Realtime, supporting emotional tone adjustment, voice cloning, and background filtering, ranking first in comprehensive voice reasoning evaluations.
Source: 通义实验室
Read on 88z.aiAlibaba Tongyi Wanxiang has released the open-source model Wan-Dancer, which generates dance videos from uploaded portrait photos and music. The model features precise beat alignment and smooth movements, overcoming long-video generation bottlenecks by supporting 720p HD output for durations ranging from 15 seconds to 3 minutes, while using a hierarchical architecture to address motion drift.
Source: Wan-Dancer
Read on 88z.aiAlipay and OPPO have partnered to enable cross-device agent interconnectivity, integrating Alipay's AI agent 'Abao' into OPPO's assistant 'Xiaobu'. Users can now use voice commands to let Xiaobu invoke Abao for nearly 200 life services, such as topping up phone credit and buying tickets, without opening apps.
Source: 支付宝
Read on 88z.aiXiaomi has open-sourced Xiaomi-Robotics-U0, a 38-billion-parameter multimodal autoregressive embodied generative foundation model that unifies four task types including scene generation, with inference efficiency improved by nearly 83 times.
Source: 小米技术
Read on 88z.aiAmap has launched ABot-World Studio, a general world model workshop that unifies interactive video generation with 3DGS scene creation for the first time. Users can create AI worlds that are real-time interactive and freely shareable using text or image inputs. The tool supports local deployment on a single 5090 GPU, maintains stable continuous inference for over an hour, and features a space-time door for cross-scene transitions.
Source: 高德
Read on 88z.aiAIsphere announced the completion of its overall Series C financing, totaling 2.98 billion yuan. The Series C+ round was led by Alibaba, with participation from over ten domestic and international institutions. Funds will be used for R&D of video generation foundation models and real-time world models, as well as global product growth. Its PixVerse platform has over 150 million global users. Earlier this year, it launched PixVerse R1, the world's first general real-time world model supporting 10
Source: 爱诗科技 AIsphere
Read on 88z.aiStepFun has released the STEPX Neo, its first AI agent phone featuring the self-developed Step AOS and personal agent Amoo. It supports natural language cross-app scheduling and cloud-edge synergy, with initial integration of major apps like Alipay and Meituan.
Source: 智东西
Read on 88z.aiAishix Technology unveiled PixVerse Game, the first real-time video game engine, at the UN AI for Good Global Summit, building full gaming experiences on real-time interactive video streams.
Source: 爱诗科技 AIsphere
Read on 88z.aiGuangzhi Shikong (Looki) announced the completion of hundreds of millions in A1 round financing, led by Yuansheng Capital, with existing shareholders such as BAI Capital increasing their stakes. The company focuses on the personal intelligent wearable sector, achieving a technical upgrade from passive response to active prediction through its PIE Active AI Engine 2.0. By combining scene understanding with long-term memory, it anticipates user needs. Its first multimodal AI hardware, Looki L1, is
Source: Founder Park
Read on 88z.aiLibTV, under Libu Libu AI, has launched LibTV Agent, the world's first professional video creation agent. It features four core capabilities: a high-baseline all-around agent, the largest professional video Skill Hub, dual-view creation, and agent-driven editing. Over 100 director-level Video Skills are available initially, covering film ads, short dramas, and self-media, with an over 80% success rate for videos under three minutes.
Source: 哩布哩布AI
Read on 88z.aiStepFun has released Step Edge, an on-device model suite for terminals like phones and cars, covering base, audio, GUI, and generation models. With native cloud-edge synergy, it achieves 0.1s low local latency and full-modal privacy protection, ranking first in 29 core benchmarks across text-vision, audio, GUI, and image generation.
Source: 阶跃StepFun
Read on 88z.aiSenseTime has officially open-sourced the SenseNova-Vision unified visual large model, natively integrating classic tasks like detection and segmentation into the large model framework, demonstrating robust capabilities in zero-shot generalization and complex scene understanding.
Source: 商汤科技SenseTime
Read on 88z.aiTencent Hunyuan has released HyOCR-1.5, the industry's first fully open-source end-to-end OCR expert large model for both training and inference. By introducing DFlash speculative decoding, the model significantly accelerates inference speeds and ranks first in end-to-end tasks on the OmniDocBench v1.6 benchmark with a score of 94.74. HyOCR-1.5 supports eight types of tasks, including document parsing, ancient character recognition, and multi-page Q&A, covering 331 languages.
Source: 腾讯混元
Read on 88z.aiThe head of OpenAI Codex engineering announced that the five-hour usage limit for Plus, Business, and Pro subscribers has been temporarily lifted for Codex and ChatGPT Work. Additionally, the GPT-5.6 Sol model will improve efficiency, allowing users to complete more tasks within the same quota.
Source: IT之家
Read on 88z.aiAgnes AI has released Agnes-2.5-Flash, a new generation text model with coding capabilities ranked among the top globally. The launch coincides with the Agnes Code desktop application, which supports local project development. Tests indicate the model can quickly locate and fix hidden bugs, independently build complex web applications, and coordinate modifications across multiple files for large-scale projects.
Source: 量子位
Read on 88z.aiOpenAI has officially released the GPT-5.6 series, featuring three models: Sol, Terra, and Luna, with API pricing ranging from $1 to $30 per million tokens. The flagship Sol model scored 53.6 in the Agents’ Last Exam benchmark, surpassing Claude Fable 5 by 13.1 points. It also achieved a record-high coding agent index of 80, while significantly reducing both cost and latency.
Source: APPSO
Read on 88z.aiMeta has officially opened the Muse Spark AI model API to developers and released the upgraded Muse Spark 1.1, marking its entry into commercial AI competition. As Meta's most powerful model for programming and agent tasks to date, it supports code generation, tool invocation, and multimodal understanding.
Source: IT之家
Read on 88z.aiOpenAI has released the new ChatGPT Work agent, powered by Codex and GPT-5.6. It supports cross-platform execution of multi-step, long-duration tasks, allowing users to hand over workflows via a single command to generate various documents and analyses.
Source: IT之家
Read on 88z.aiAnt Lingbo Tech has released LingBot-VA 2.0, the industry's first embodied-native world action model. Built on a zero-based pre-trained autoregressive architecture, it integrates four core designs: semantic visual-action tokenizer, causal pre-training, MoE architecture, and asynchronous inference. It supports 150Hz real-time inference on a single card, enabling robots with general control capabilities to reason and act simultaneously.
Source: 蚂蚁灵波科技
Read on 88z.aiKuaishou KwaiKAT has released its flagship Agentic Coding model, KAT-Coder-Pro V2.5, achieving breakthroughs in long-horizon engineering, general agentic capabilities, and large-scale reinforcement learning. By constructing over 100,000 runnable repositories via AutoBuilder and employing multi-framework RL with Mixture of Experts techniques, the model leads in benchmarks such as SWE-Bench Pro.
Source: KwaiKAT
Read on 88z.aiByteDance's Seed team has released Seedream 5.0 Pro, significantly enhancing text-image alignment, text rendering, and aesthetic quality. The model achieves breakthroughs in visualizing complex information, interactive precise editing, realistic image texture restoration, and native multilingual generation. It supports converting data and dense text into professional layouts and allows for refined control via point-and-click selection.
Source: 字节跳动Seed
Read on 88z.aiOpenAI has released GPT-Live, a new voice model featuring real-time simultaneous interpretation, full-duplex conversation, and customizable reasoning intensity. It employs deep task delegation to handle complex background tasks while maintaining smooth interaction, and supports real-time visual card displays.
Source: 量子位
Read on 88z.aiSpaceXAI has released its flagship model, Grok 4.5, featuring a 1.5T parameter MoE architecture and joint training with Cursor. It rivals Claude Opus in performance, achieving 80 TPS inference speed and surpassing Opus 4.8 in programming on DeepSWE Bench, with over 4x better token efficiency. It currently supports a 500k context window, expanding to 1 million next week, with a 2T parameter version coming later this month.
Source: 量子位
Read on 88z.aiAnt Lingbo Technology has open-sourced LingBot-Video, the world's first Mixture-of-Experts (MoE) foundation model for embodied AI video generation. With 30 billion total parameters and only 3 billion activated during inference, it offers triple the efficiency of dense architectures. The model leverages a 70,000-hour dataset and multi-dimensional reinforcement learning rewards to improve physical plausibility and task completion.
Source: 蚂蚁灵波科技
Read on 88z.aiMistral AI has released Robostral Navigate, an embodied AI navigation model that enables autonomous navigation in complex environments using only a single RGB camera. The 8-billion parameter model performs well in unknown environment tests and significantly reduces training time through technical optimizations, supporting continuous self-improvement.
Source: Navigate
Read on 88z.aiAnthropic announced that the Claude Cowork task agent has officially expanded to web and mobile platforms, supporting cross-device session continuity and file operations. Users can now initiate tasks on desktop and monitor progress on mobile, with scheduled tasks capable of running offline in the background. Claude sends confirmation requests to users' phones for key decisions. The feature is currently in beta and initially available to Max subscribers.
Source: IT之家
Read on 88z.aiMeta's Superintelligence Lab has released its first self-developed AI image generation model, Muse Image. The model supports text-to-image creation, editing, and QR code generation, ranking second on the Arena AI leaderboard while providing various free AI effects for Instagram Stories.
Source: IT之家
Read on 88z.aiAnt Lingbo Technology has released the spatial perception model LingBot-Depth 2.0, with training data significantly expanded to 150 million. It performed excellently in depth completion benchmarks. The company also open-sourced the visual foundation model LingBot-Vision and achieved certification cooperation with Orbbec.
Source: 蚂蚁灵波科技
Read on 88z.aiBAAI has released WuJie·RoboBrain Orca, a multimodal representation world model. The model utilizes 125,000 hours of video for unconscious learning and combines this with 160 million event annotations for conscious learning to construct a unified latent representation space of the world. It outperforms baseline models of similar scale in downstream tasks such as text generation, image prediction, and embodied action generation.
Source: 智源研究院
Read on 88z.aiMomenta listed on the HKEX under code 6880.HK. Its opening price rose over 6% on the first day, pushing the market cap above HK$70 billion. The company raised about HK$6.8 billion, with public offering oversubscribed 414 times and participation from 14 cornerstone investors including GIC and Fidelity.
Source: 财联社
Read on 88z.aiByteDance's Seed team has released EdgeBench to evaluate the continuous learning capabilities of AI Agents in real-world environments. The benchmark includes 134 real tasks across six major domains, supporting extended continuous operation. Research indicates that Agent learning trajectories follow a high-precision log-sigmoid curve, with leading models doubling their learning speed every three months.
Source: 字节跳动Seed
Read on 88z.aiThe Alipay AI Open Platform has officially launched and opened for invited testing, offering a one-stop AI upgrade solution for merchants and developers. Existing applications can be quickly transformed into MCPs, Skills, or Agents, connecting with a billion users via the Abao assistant and covering multiple terminals including phones, car systems, and AI glasses.
Source: 支付宝
Read on 88z.aiTencent's AI-generated application app 'Toast' has officially launched on Apple iOS, following the Android release last month. Positioned as an exploratory atmospheric programming tool, it enables users without coding skills to turn ideas into functional apps, creating a closed-loop ecosystem from creation to distribution.
Source: IT之家
Read on 88z.aiShanghai AI Lab has released the open-source research agent workbench InternAgentS, designed specifically for the AI for Science domain. The platform supports local deployment and multi-model adaptation, integrating full-process functions such as paper reading, experiment analysis, code iteration, and scientific writing. It is compatible with domestic and private models like DeepSeek, Qwen, and Kimi, connecting to over 3,600 research tools and skills.
Source: 书生Intern
Read on 88z.aiAlibaba Cloud has launched Qoder Enterprise Edition, featuring the QMind personal cloud knowledge base for multi-source data integration and cross-dimensional knowledge sharing. The new version introduces a pooled Credits payment model allowing administrators to dynamically allocate quotas, along with five-layer security protection covering transmission encryption to AI runtime, certified under ISO 27001.
Source: 阿里云
Read on 88z.aiMeituan has officially open-sourced LongCat-2.0, a trillion-parameter large model designed specifically for Agentic Coding tasks. With 1.6T total parameters and 48B average active parameters, the model introduces innovative sparse attention and N-gram Embedding techniques to enhance long-context processing and code execution performance. It has completed inference validation on a domestic computing cluster of 50,000 cards. The release includes multi-precision versions (BF16/FP8/INT8) and optimiz
Source: 龙猫LongCat
Read on 88z.aiTencent Hunyuan has officially released the Hy3 large model. By improving training data quality and scaling reinforcement learning compute, it achieves significant progress in reasoning, agents, and long-context tasks. Its performance rivals flagship models with 2-5 times more parameters, while substantially reducing hallucination rates, common sense errors, and multi-turn issue rates.
Source: 腾讯混元
Read on 88z.aiAlibaba has launched the upgraded Fun-ASR-Realtime, a real-time speech recognition model with sub-second initial latency and accuracy comparable to offline systems. It supports 30 languages and 16 dialects, featuring context awareness and self-correction capabilities, outperforming competitors in dialect recognition tests.
Source: 千问大模型
Read on 88z.aiSenseTime has officially released SenseNova U1 Infographic Enhanced Version V2. Built on the 8B-MoT architecture, the model offers lighter and more open capabilities for generating high-information-density images. The new version significantly improves small text clarity, complex layout structures, and visual texture, aiming to bring AI-generated infographics to professional designer standards.
Source: 商汤科技SenseTime
Read on 88z.aiShengshu Technology unveiled the Vidu S1 real-time interactive model at the 2026 Global Digital Economy Conference. The model supports voice-controlled video generation and unlimited continuous interaction, featuring 540P resolution and 25 FPS. Users can create personalized characters such as humans, anime figures, or pets by uploading a single image without additional modeling training.
Source: 生数科技
Read on 88z.aiMistral AI has released the open-source model Leanstral 1.5, optimized for automated theorem proving in Lean 4. Featuring a 119B parameter MoE architecture, it supports ultra-long context and multimodal inputs, achieving top performance on math benchmarks, and is now available for download and online use.
Read on 88z.aiGoogle has released TabFM, an open-source zero-shot tabular foundation model that handles structured data with mixed features without fine-tuning, achieving strong results across multiple benchmarks.
Source: TabFM
Read on 88z.aiNVIDIA has open-sourced the Nemotron-Labs-TwoTower model, utilizing a 60B parameter dual-tower architecture to address token generation bottlenecks. This discrete diffusion language model combines a 30B autoregressive context tower with a 30B diffusion denoising tower, enhancing efficiency through cross-attention collaboration.
Source: IT之家
Read on 88z.aiAlibaba has internally announced a complete ban on Anthropic's Claude product line, including models like Sonnet, Opus, and Fable, as well as agent products such as Claude Code. The trigger was the exposure of security risks involving backdoors in Claude Code, which has included a hidden trojan system since version 2.1.91 that marks Chinese users by detecting local time zones and proxy keywords.
Source: 智东西
Read on 88z.aiKling AI, under Kuaishou, plans to close a $3 billion financing round with Tencent participating, setting a post-money valuation of $18 billion. The company expects to launch a Hong Kong IPO within a year, using funds for computing power and talent. Its annualized revenue exceeds $300 million, mainly from overseas markets.
Source: 智东西
Read on 88z.aiAlipay's AI agent 'Abao' has ended its internal testing and is now open for public beta on iOS and Android without invite codes. Users can access it by searching in app stores or within the Alipay app. The agent integrates 72 life service skills, supporting fuzzy spoken language understanding and full-scenario interaction.
Source: 支付宝
Read on 88z.aiKunlun Wanwei's TianGong 3.2 has upgraded with Skywork Tags, enabling Skywork Agent integration into instant messaging tools such as Slack, Feishu, DingTalk, Discord, and Telegram. It resides in work group chats as a 'team member.' Unlike traditional models that require users to migrate context to separate platforms, Skywork Tags operates directly within existing team communication environments, eliminating the need to export chat logs or switch windows.
Source: 昆仑万维集团
Read on 88z.aixAI has released the beta version of Voice Agent Builder, a zero-code platform built on the Grok Voice end-to-end speech model. Users can configure agents in two minutes using natural language, with support for SIP access, real-time knowledge base retrieval, and various tool integrations, along with brand cloning and full observability.
Source: Builder
Read on 88z.aiYuankong AI has officially launched the desktop version of ChatExcel Ultra, positioned as an enterprise data office tool. This version emphasizes local data processing, keeping files entirely on local machines or intranets without passing through external servers, thereby meeting compliance requirements for sensitive sectors like finance, healthcare, and government. Additionally, the product supports direct database connections and cross-database joint analysis, allowing users to perform complex
Source: 元空AI
Read on 88z.aiMeta has officially open-sourced Astryx, a design system built on React and StyleX. Currently in beta, the system underwent eight years of internal refinement and supports over 13,000 applications including Facebook and Instagram.
Source: Astryx
Read on 88z.aiAnthropic has released Claude Sonnet 5, its most capable Sonnet model for agents, which can autonomously execute complex tasks with performance nearing Opus 4.8 at a lower price, now available to all subscription tiers.
Source: APPSO
Read on 88z.aiGoogle has released Nano Banana 2 Lite, a new lightweight AI image generation model focused on speed and batch processing. It generates images in just four seconds at a cost of $0.034 per thousand. Available on Google AI Studio, Gemini API, and enterprise platforms, it is designed for high-frequency content creation and will replace the original Nano Banana.
Source: IT之家
Read on 88z.aiAnthropic has released Claude Science, an AI work platform for researchers that offers an integrated environment for literature analysis, multi-step research execution, and result generation, ensuring outputs are auditable and reproducible.
Source: IT之家
Read on 88z.aiTransformer-specific AI chip startup Etched has completed the tape-out of its A0 chip, securing $800 million in funding and $1 billion in customer orders. Founded by a Harvard dropout, the company is backed by AI luminaries including Karpathy, Li Fei-Fei, and Hinton. Etched has launched its first rack product, utilizing Low Voltage Inference and Cluster-Scale Memory technologies optimized for trillion-parameter MoE model inference, with shipments planned for this summer.
Source: 量子位
Read on 88z.aiOpenAI announced a collaboration with Work Louder to launch a Codex-branded keyboard on July 15. The device is built on the Creator Micro 2 and optimized for AI development.
Source: IT之家
Read on 88z.aiMeituan has officially open-sourced LongCat-2.0, a large model with 1.6 trillion total parameters and 48 billion average activated parameters. It is the first model in the industry to complete full-process training and inference on a domestic computing cluster of 50,000 cards, natively supporting 1M long context and optimized for Agentic Coding tasks.
Source: 龙猫LongCat
Read on 88z.aiTencent Video has launched WorkRally, an industrial-grade AI platform for producing comic dramas. It targets 2D/3D animation and AI-generated human-like dramas, covering the entire workflow from script analysis to asset management. The platform features an S+ level film and animation skill library, supports expert-level Agent generation and intelligent pipeline scheduling, and includes capabilities for character consistency management and error detection.
Source: WorkRally
Read on 88z.aiTsinghua-affiliated startup Liqing Intelligence has completed a seed funding round worth hundreds of millions, with participation from Shunwei Capital, HongShan China, and Gaoling Ventures. Founded by Tsinghua AI Institute assistant professor Li Yiming, the team has an average age of 23. The company positions itself as a Physical AI infrastructure provider, developing proprietary data pipelines and differentiable physics engines to achieve a Real-to-Sim-Real loop, requiring only 1% of real-world
Source: 智能涌现
Read on 88z.aiHuawei has announced the open-source release of its 92-billion parameter openPangu-2.0-Flash model. Built on Ascend native technology, it aims to provide foundational support for agent applications. Model weights, inference code, and key operators are now available on open-source platforms, with the Pro version expected in July and additional components rolling out in the second half of the year.
Source: 华为
Read on 88z.aiGlobal fintech platform Airwallex announced the completion of a $320 million Series H financing round, reaching a post-money valuation of $11 billion. Led by Addition, the funds will support AI product development, global compliance, and team expansion, alongside the launch of two new AI products.
Source: Z Potentials
Read on 88z.aiOpenAI has released a limited preview of the GPT-5.6 series, featuring Sol, Terra, and Luna models. Sol enhances professional capabilities with new deep reasoning and multi-agent modes; Terra offers equivalent performance at half the price; Luna focuses on being lightweight and fast. Due to US government requirements, the series shifted from planned open access to a limited preview.
Source: 机器之心
Read on 88z.aiDeepSeek and Peking University have jointly released DSpark, a large model inference acceleration framework now deployed in the DeepSeek-V4 production system. While maintaining total throughput, DSpark significantly boosts single-user generation speed: DeepSeek-V4-Flash sees a 60%–85% increase, and DeepSeek-V4-Pro a 57%–78% rise. It utilizes a semi-autoregressive architecture with confidence-based scheduling verification to balance draft generation efficiency and text coherence, dynamically adju
Source: APPSO
Read on 88z.aiAmap is internally testing a Vibe Coding product named 'Daima'. The platform focuses on zero-threshold app building driven by natural language, targeting WeChat Mini Programs and iOS native scenarios. It enables users to quickly generate applications that are ready for launch and functional on real devices.
Source: IT之家
Read on 88z.aiAlibaba's Tongyi Qianwen team has launched a standalone AI voice input method called 'Qianwen Input', emphasizing casual speech and polished writing. Built on the Qianwen large model's CosyVoice capabilities, it converts speech to text while understanding semantics, removing filler words, and correcting slips to structure spoken language into written text. It supports speeds up to 300 characters per minute, recognition of nine dialects, and mixed Chinese-English speech.
Source: 千问输入法
Read on 88z.aiShengshu Technology has launched Vidy, an AI companion app positioned as an 'AI Vibe Buddy' that supports text, voice, and real-time calls. Featuring characters like Humi and a relationship growth memory system, it leverages the self-developed U-ViT architecture and Vidu video large model to plan for real-time AI facial expression and motion generation, upgrading companionship to face-to-face video interactions.
Source: Vidy
Read on 88z.aiX Variable completed four consecutive funding rounds within two months, with its post-money valuation exceeding 20 billion yuan, making it the first embodied brain company in the Greater Bay Area to reach this milestone. Meituan, Alibaba, ByteDance, and Xiaomi each led a round, with Xiaomi investing in three consecutive rounds. Over thirty institutions, including 58 Group, Chery, China Mobile, and Sequoia China, participated.
Source: 量子位
Read on 88z.aiDeepReinforce has released the open-source Agentic programming model Ornith-1.0, built on Gemma 4 and Qwen 3.5 bases with four sizes ranging from 9B to 397B. The series performs strongly on Terminal-Bench 2.1 and SWE-Bench Verified benchmarks, with the flagship 397B version surpassing Claude Opus 4.7 and the 35B MoE version outperforming Qwen 3.5-397B.
Source: Ornith-1.0
Read on 88z.aiTencent Hunyuan has officially open-sourced PhoneBuddy, a mobile agent model with 4 billion parameters. Alongside the release, the team published five academic papers, including PhoneWorld and PhoneHarness, establishing a comprehensive technical framework covering environment construction, training, execution, and privacy security. The model's core innovation lies in its Real+Mock hybrid reinforcement learning strategy, which combines operations in real apps with a self-built PhoneWorld simulati
Source: 量子位
Read on 88z.aiMistral AI has released Mistral OCR 4, a new document recognition model supporting 170 languages. It scored 93.07 on the OmniDocBench benchmark, outperforming some competitors in human preference. The model provides bounding boxes, region classification, and confidence scores, making it suitable for downstream tasks like RAG.
Source: IT之家
Read on 88z.aiOpenAI has unveiled Jalapeño, its first in-house inference chip, developed with Broadcom and Celestica under the leadership of a former Google TPU core member. Tailored for models like ChatGPT, it set a record for fastest ASIC tape-out and is scheduled for deployment by late 2026 with superior energy efficiency.
Source: 量子位
Read on 88z.aiHuawei has announced that the HarmonyOS 'Lobster' Xiaoyi Claw is now available for all models. Devices running HarmonyOS 5.0 or later can access it by updating the Xiaoyi App to the specified version. The updated service packages include a 49 yuan trial pack with Auto-Model mode and a 199 yuan standard pack offering choice among four models, including openPangu-2.0-Pro.
Source: IT之家
Read on 88z.aiAlibaba Cloud has released CosyVoice, a new AI voice input method powered by the Qwen large model. It offers semantic-level understanding and text regeneration, supporting real-time transcription, redundancy filtering, and multi-dialect recognition to surpass traditional acoustic transcription.
Source: CosyVoice
Read on 88z.aiAlibaba Cloud has launched 'Xiangwei', an AI-powered platform for interactive story creation and reading. It supports three modes: starting from scratch, continuing existing text, and planning chapters for long novels, catering to web novels, scripts, and children's stories. During reading, AI generates background images in real time, allowing readers to choose plot directions or input custom text at key points, with the platform instantly continuing the story while maintaining character consist
Source: 向尾
Read on 88z.aiNetEase Youdao has officially open-sourced the Confucius4-TTS speech synthesis model. The model achieves breakthroughs in zero-shot cloning, multi-language accent-free synthesis, and emotion transfer, utilizing an end-to-end architecture that supports local offline deployment.
Source: 网易有道
Read on 88z.aiDoubao Professional Edition is officially launched, introducing a new office task mode based on the Doubao 2.1 series of large models. It supports local computer operations, browser integration, Skills invocation, scheduled tasks, and built-in Office suites, enabling professional work such as document processing, application development, and data analysis. Subscription features three-tier pricing: Standard at 68 yuan/month, Enhanced at 200 yuan/month, and Premium at 500 yuan/month; verified univ
Source: 豆包
Read on 88z.aiVolcengine has released the Doubao Audio Generation Model 1.0, which supports text or audio reference inputs to generate complete audio end-to-end, including multi-character dialogues, emotions, and background music, while maintaining consistent voice tones over long durations.
Source: 火山引擎
Read on 88z.aiAnthropic has released Claude Tag, transforming its AI assistant into a virtual colleague within Slack channels. The tool supports task assignment via mentions, automatically breaks down and executes workflows, and features team relay collaboration, organizational memory retention, proactive monitoring alerts, and asynchronous long-task processing to enhance enterprise collaboration efficiency.
Source: APPSO
Read on 88z.aiAlibaba Tongyi has launched Qwen-AgentWorld, the first native language world model, supporting agent interaction simulation across seven domains. Trained on tens of millions of real-world trajectories, it outperforms top models like GPT-5.4 and Claude Opus 4.8 in proprietary benchmarks.
Source: 千问大模型
Read on 88z.aiTRAE Work has launched a new Design mode, available simultaneously on desktop and web. Users can generate design proposals through natural language conversations, modify them in real-time on the canvas, and create interactive prototypes. The mode supports importing existing design assets from platforms like Figma and allows one-click export of designs to Code mode, enabling a seamless workflow from design to development.
Source: TRAE.ai
Read on 88z.aiBaichuan Intelligence and Tsinghua University jointly released the medical-enhanced large model Baichuan-M4, which ranked first globally in three HealthBench categories with an industry-low hallucination rate of 3.3%.
Source: 百川智能
Read on 88z.aiAutonomous driving firm Momenta is pursuing a Hong Kong listing, positioning itself as a builder of physical AI foundation models. Its R7 world model is mass-produced, leveraging massive data and a three-layer architecture to understand the physical world. Revenue has tripled in three years, with licensing income surging 42-fold as the business model shifts to scalable licensing fees.
Source: 量子位
Read on 88z.aiShadow眸 Technology announced the completion of a new funding round worth hundreds of millions, led by Cathay Capital and Shanghai State-owned Capital Venture Guide, with existing shareholders continuing to participate. The company simultaneously released the Hyper3D Rodin Gen-2.5 model, introducing an LLM think-then-generate mechanism with five levels of thinking depth, capable of generating million-face models in as fast as 4 seconds, and launched the world's first native 3D texture model with
Source: 智能涌现
Read on 88z.aiByteDance has released the Seed2.1 series of models, positioned as new agents for real-world productivity scenarios. The model shows significant improvements in general agent capabilities, end-to-end code delivery, and multimodal understanding. It leads in multiple benchmarks and surpasses Claude Opus 4.6 in coding crowd-testing win rates.
Source: 字节跳动Seed
Read on 88z.aiTRAE Work China has integrated the Doubao 2.1 model and offers it for free, featuring enhanced Agent and multimodal capabilities. TRAE IDE is currently available only to specific paid users.
Source: TRAE.ai
Read on 88z.aiByteDance officially released the Doubao video generation model Seedance 2.5 at the Volcano Engine FORCE conference. Currently in the final stages of beta testing, the model is scheduled for launch in early July, featuring breakthroughs in video duration, multimodal input, and controllable editing.
Source: APPSO
Read on 88z.aiBaidu has released the open-source OCR model Unlimited OCR, specializing in long document parsing. It achieved a new state-of-the-art comprehensive score of 93.23% on the OmniDocBench benchmark, surpassing DeepSeek OCR. The core innovation is the Reference Sliding Window Attention mechanism, which uses a soft forgetting strategy to maintain constant KV Cache, ensuring inference speed does not drop as document length increases, with TPS improving by approximately 35% at 6000 tokens.
Source: 量子位
Read on 88z.aiUS AI inference chip unicorn Groq announced the completion of a $650 million financing round to accelerate global AI inference cloud expansion, targeting 200 megawatts of computing power by 2027. The company previously reached a massive licensing agreement with Nvidia, with its core team joining the latter. It now focuses on inference cloud operations, running 13 data centers and serving over 5 million developers, while appointing a new leadership team.
Source: 智东西
Read on 88z.aiWeChat has launched a limited beta test for its native AI assistant, 'Xiao Wei,' accessible via a green eye icon in the top-left corner of the main interface. The assistant supports text and voice interactions to perform native functions such as adjusting settings, sending messages, launching mini-programs, and generating images. It also features the ability to create mini-programs from a single sentence. The core model is WeChat's self-developed WeLM, with some responses powered by DeepSeek.
Source: 财联社
Read on 88z.aiZhipu AI's total market capitalization on the Hong Kong stock exchange has surpassed 1 trillion HKD for the first time, rising over 2000% this year to reach 1.27 trillion HKD. Its open-source GLM-5.2 model ranked first among usable models in frontend development evaluations.
Source: IT之家
Read on 88z.aiTencent QQ Mail has launched Agently Mail, a standalone email service designed for AI Agents. Users can obtain a dedicated address via WeChat authorization, ensuring isolation from personal inboxes while offering full email management capabilities. The service is now available on Tencent SkillHub.
Source: Mail
Read on 88z.aiDuiyou platform officially launches Alibaba's HappyHorse 1.1 video generation model. The new version achieves systematic optimization in five dimensions: dynamic expression, consistency, visual quality, instruction following, and camera language. Users can experience it directly via the official website.
Source: 堆友
Read on 88z.aiAnthropic announced that Claude Code now supports the Artifacts feature, which can real-time generate shareable interactive web pages from session content such as code modifications and troubleshooting, facilitating team progress synchronization. This feature is automatically generated based on full session context, with pages continuously refreshing as work progresses. It is private by default and limited to sharing within the organization. Currently, Artifacts are available in Beta to Team and
Source: APPSO
Read on 88z.aiAlibaba Cloud launches HappyOyster 1.0, an open-world model enabling real-time interaction. It generates explorable digital worlds via natural language and supports multimodal input with real-time responsiveness.
Source: IT之家
Read on 88z.aiEVOKEN has completed a Series B+ financing round of nearly $300 million, reaching a post-money valuation exceeding $2 billion. The round was co-led by Granite Asia and Shunwei Capital, with continued participation from existing shareholders such as Gaorong Ventures and Ant Group. The company operates three major platforms: LibLibAI, LibTV, and Xingliu. AI video platform LibTV saw revenue grow more than 13-fold in its first two months, contributing to a year-over-year group revenue increase of ov
Source: 演语 EVOKEN
Read on 88z.aiDeepSeek multimodal researcher Xiaokang Chen announced that the DeepSeek image recognition mode is now fully available on both web and app platforms. Alongside fast and expert modes, this feature allows users to upload images for deep understanding by DeepSeek, offering capabilities far beyond simple text extraction.
Source: IT之家
Read on 88z.aiVolcengine has officially launched the API service for Doubao Real-time Voice Model 3.0 (Seeduplex) and opened it for invited testing. As a native full-duplex end-to-end voice large model, it features precise adherence, interference resistance, and dynamic stop detection. It supports tool invocation during real-time conversations, enabling listening, speaking, and task execution simultaneously. Stop detection latency is reduced by approximately 250ms, and the interruption rate in complex scenari
Source: 火山引擎
Read on 88z.aiOpenAI is gradually rolling out a new scheduled tasks feature for ChatGPT, replacing the existing Pulse functionality. This update aims to improve speed and stability, enabling users to send reminders, handle periodic work, and monitor specific items. The new interface allows unified task management via a dedicated sidebar page, supporting both precise times and broad time windows. Monitoring tasks can perform web searches and read app data, but are limited to running once per hour and may pause
Source: IT之家
Read on 88z.aiSpaceX has announced an all-stock acquisition of Anysphere, the parent company of AI coding tool Cursor, for $60 billion. The deal is expected to close in Q3 2026, making Cursor a wholly-owned subsidiary of SpaceX.
Source: 机器之心
Read on 88z.aiDeepSeek has completed its first external funding round, raising over 50 billion yuan with a valuation exceeding 330 billion yuan, setting a record for the largest single financing in China's AI industry. The deal features a unique structure where funds are directed into a limited partnership managed by Liang Wenfeng. External investors face a five-year lock-up period and hold no voting rights, except for the National AI Industry Investment Fund which invested 1 billion yuan directly. Liang Wenf
Source: 机器之心
Read on 88z.aiZhipu AI has launched and open-sourced the GLM-5.2 model, which ranks first among globally available models in Code Arena front-end evaluations. Designed for long-horizon tasks, it supports a 1M lossless context window and performs close to Claude Opus 4.8. The model introduces effort level control and an IndexShare architecture to reduce FLOPs, with native adaptation for domestic computing power such as Huawei Ascend.
Source: 智谱
Read on 88z.aiXiaomi has officially launched Xiaomi MiMo Claw, featuring the MiMo-V2.5-Pro flagship model. It natively supports the OpenClaw framework and MCP protocol, handling million-level long contexts. The product integrates deeply with the Kingsoft Office ecosystem, enabling end-to-end generation and editing for Word, Excel, PPT, and PDF documents. It uses a cloud-based, deployment-free architecture to provide 24/7 dedicated Agent services.
Source: Xiaomi MiMo
Read on 88z.aiWeChat Pay has launched the AI Exclusive Card, with WorkBuddy being the first to integrate it. Users can have AI handle restaurant searches, coupon collection, group buying, and payments via simple voice commands. Fund security is ensured through dedicated funds and transaction-by-transaction confirmation, requiring mobile authorization for each payment. Currently, it supports only WorkBuddy Mac version 5.1.1 and above, accessible via the Meituan Life Assistant in the expert page.
Source: 腾讯
Read on 88z.aiMiniMax has released MiniMax M3, an open-source native multimodal flagship model. With 428B total parameters and 23B active parameters, it utilizes a proprietary MSA sparse attention architecture to reduce long-context computational costs. As the first open-source model trained with multimodal mixing from step zero, M3 ranked first among open-source models on benchmarks including Artificial Analysis Comprehensive Intelligence Index, GDPval-AA, and Code Arena WebDev, earning praise from industry
Source: MiniMax 稀宇科技
Read on 88z.aiAlipay has launched the AI Version of Alipay, allowing users to access a clean chat interface by swiping right. Users can handle over ten thousand services in government affairs, travel, and daily life via voice or text commands, realizing chat-as-a-service. Fund-related actions require user confirmation for security. The AI version is currently in invite-only testing, allows free switching with the classic version, and opens its service ecosystem to promote inclusive AI for one billion users.
Source: 支付宝
Read on 88z.aiAlibaba Tongyi has released the Qwen-Robot Suite, integrating navigation, manipulation, and world prediction models. It aligns vision and action via a language-first interface, achieving SOTA results in over ten benchmarks.
Source: 千问大模型
Read on 88z.aiJimeng AI has launched the Seedance 2.0 Mini video generation model, focusing on extreme cost-effectiveness with costs as low as 0.16 yuan per second for 720P resolution. The model is open to creators, who can access it via the web platform at introductory prices for affordable AI video creation.
Source: 即梦AI
Read on 88z.aiVolcano Engine's Volcano Ark Experience Center has launched the Seedance 2.0 mini video generation model, with API services opening soon. Positioned for high cost-performance, it prices image-to-video at 0.023 yuan per thousand tokens and video-to-video at 0.014 yuan. At 720P resolution, the cost is approximately 0.5 yuan per second, a 50% reduction from the previous version, catering to broad creation and scaled production needs.
Source: 第一财经
Read on 88z.aiZhipu AI has released the GLM-5.2 model, which supports a 1M context window and is positioned as the leading domestic coding model. It will be available to all GLM Coding Plan users, with its API launching this week and the model open-sourced under the MIT license.
Source: 智谱
Read on 88z.aiThe Kimi API Open Platform has launched the Kimi K2.7 Code High-Speed version, offering output speeds 5-6 times faster than the standard version at twice the price. Compared to K2.6, this model shows significant improvements in long-context programming and token efficiency, reducing average token consumption by 30% and boosting benchmark performance by 11%-31.5%.
Source: Kimi开放平台
Read on 88z.aiDaxiao Robotics has completed its Angel+ funding round, bringing its cumulative financing in the first half of the year to hundreds of millions of dollars. The round saw participation from investors such as Fortune Capital and Shenzhen Capital Group. The funds will be directed toward the development of world models and integrated hardware-software solutions. Its flagship product, Kaiwu World Model 3.0, is the world's first open-source mass-produced world model, ranking first in multiple authorit
Source: 商汤科技SenseTime
Read on 88z.aiBaidu PaddlePaddle has released PP-OCRv6, offering Tiny, Small, and Medium variants for the first time. Detection and recognition accuracy have improved by 4.9% and 5.1% respectively over the previous version, with CPU inference speed up to 5.2 times faster. The new model supports 50 languages, adds industrial scenarios like circuit boards and digital tubes, and features enhanced robustness. It is now available on the official website and open-sourced on GitHub and HuggingFace.
Source: 飞桨 PaddlePaddle
Read on 88z.aiMoonshot AI has released the open-source programming model Kimi K2.7 Code, which shows significant improvements over K2.6 in long-context processing, instruction following, and autonomous Agent execution. The model achieves an 11% to 31.5% increase in code benchmark scores and reduces token consumption by 30%, while maintaining the same API pricing as its predecessor. Enabling thinking mode is essential to unlock its optimal performance.
Source: 月之暗面 Kimi
Read on 88z.aiMiniMax has launched MiniMax Hub, a desktop multimodal creative Agent platform that integrates copy, image, video, and audio Agents onto a single canvas, supporting parallel workflows for one-stop creation from concept to final output.
Source: 海螺AI
Read on 88z.aiOpenAI has announced the acquisition of startup Ona to provide secure, pre-configured cloud environments for its coding assistant, Codex. This integration aims to enhance Codex's ability to handle long-running tasks and enable users to safely deploy AI agents into production settings. It also strengthens enterprise control over infrastructure, data assets, and security boundaries. Financial terms were not disclosed, and the Ona team will join OpenAI to work on Codex development.
Source: IT之家
Read on 88z.aiNous Research has launched Profile Builder for Hermes Agent, converting command-line configuration into a web-based visual workflow that allows users to create agents in five steps.
Source: IT之家
Read on 88z.aiVisa and OpenAI have announced a strategic partnership to drive agent commercialization and explore new transaction models. Visa is integrating its global payment network, tokenization, authorization, identity verification, and fraud detection technologies into OpenAI products to ensure the security of AI-initiated transactions while retaining final user decision-making authority. The two parties are also exploring the integration of payment capabilities into developer tools.
Source: IT之家
Read on 88z.aiXiaomi's large model team has launched MiMo Code, an open-source terminal AI coding assistant. It features the time-limited free MiMo-V2.5 model and supports mainstream models like DeepSeek and Kimi. Key capabilities include a persistent memory system, Compose mode for full-process automated development, memory organization commands, and voice control. In SWE-Bench Pro and Terminal Bench 2 benchmarks, it outperforms Claude Code under identical model conditions.
Source: Xiaomi MiMo
Read on 88z.aiHiDream Future has released the commercial image generation model HiDream-O1-Image-1.5. Ranking third globally and first in China on the Artificial Analysis text-to-image leaderboard with an ELO of 1265, it trails only OpenAI's GPT Image series while surpassing major models from Google, NVIDIA, and ByteDance. Built on the native multimodal UiT architecture, it excels in semantic adherence, text rendering, complex layout, and multi-subject consistency.
Source: HiDream智象未来
Read on 88z.aiAlibaba Meoo has released Meoo CLI, an open-source command line tool designed to streamline the deployment of local AI programming projects. The tool enables local agents such as Claude Code, Codex, and Cursor to invoke cloud capabilities via natural language commands, automatically handling database connections, user authentication, file storage, and model calls. This allows developers to transform local demos into publicly accessible online applications within ten minutes.
Source: 阿里云
Read on 88z.aiiFlytek has launched SpaceMind, a smart space Agentic architecture that uses millimeter-wave radar as its core sensor and combines a dual-routing architecture with edge-cloud collaborative models to push smart homes into the L2.5 era of proactive intelligence. The architecture supports 5-centimeter-level positioning, breath-level micro-motion detection, and response times as low as 700 milliseconds, featuring persistent memory, multi-Agent collaboration, and proactive services based on user habi
Source: 讯飞开放平台
Read on 88z.aiAniShort, the AI short drama collaboration platform under Badianba Digital, has completed nearly 100 million yuan in financing led by Beijing Taizhonghe. The funds will support R&D and creator ecosystem development, aiming to address high costs and long production cycles by streamlining the entire workflow.
Source: 量子位
Read on 88z.aiGUI Agent execution platform Core-Mate has completed a funding round of tens of millions of yuan, exclusively invested by Dunhong Assets. Core team members are from ByteDance, having served as product managers for Douyin and Doubao. The platform enables AI to handle real operations in web pages and apps by integrating vision, control trees, and page states via the open-source OpenGUI framework for stable screen semantic understanding.
Source: Founder Park
Read on 88z.aiAnthropic has released its most powerful model, Claude Fable 5, which outperforms previous generations across the board and leads significantly in benchmarks like SWE-Bench Pro. The model offers higher token efficiency, capable of migrating a 50-million-line codebase in one day. Additionally, Anthropic launched the unguarded version, Claude Mythos 5, available exclusively to Glasswing partners. For safety, Claude Fable 5 includes guardrails that downgrade sensitive requests to Claude Opus 4.8.
Source: 机器之心
Read on 88z.aiByteDance's TRAE announces that TRAE SOLO has officially upgraded to TRAE Work, expanding its brand meaning from The Real AI Engineer to The Real AI Enabler. The new version targets broader real-world work scenarios with Work and Code modes, covers PC, mobile, and Web platforms, supports parallel multitasking, and completed integration with Feishu in May.
Source: Work
Read on 88z.aiAnt Group has launched QMuse, an AI app generation platform that enables zero-code creation of runnable applications via natural language. Currently in beta and requiring an invite code, the platform aims to lower development barriers while ensuring security.
Source: QMuse
Read on 88z.aiGoogle has released Gemini 3.5 Live Translate, supporting real-time translation for over 70 languages while preserving original tone and speed with only a few seconds of latency. Developers can access the API preview, enterprises can use it in Google Meet, and general users can try it via the Google Translate app.
Source: IT之家
Read on 88z.aiAlibaba's Qwen app has launched China's first full-cycle college application agent. Leveraging a large model and eight years of Quark data, it offers free one-stop services including application calendars, reports, and Q&A for students.
Source: 千问APP
Read on 88z.aiAmap has released ABot-Earth 0.5, the first 3D-native urban world model. It supports multimodal inputs and generates high-fidelity scenes quickly on consumer-grade GPUs, covering numerous countries and regions worldwide.
Source: 高德技术
Read on 88z.aiXiaomi, in collaboration with TileRT, has launched the MiMo-V2.5-Pro UltraSpeed mode, achieving a generation speed of over 1000 tokens/s for trillion-parameter models on general-purpose GPUs. Priced at three times the standard version, it offers approximately tenfold faster output and is available exclusively via API. Access is limited and application-based, prioritizing enterprises and professional developers.
Source: IT之家
Read on 88z.aiJingying Technology has completed tens of millions of USD in Series A and A+ financing, with investors including Wang Huiwen's family office and Ant Group. The company also announced that Wang Minjie, former Chief Applied Scientist at AWS, has joined as Chief Scientist. Positioned as the first Agent-native company in the content industry, it focuses on AI short dramas to build a reinforcement learning environment for creator Agents that is accessible and self-evolving, driven by real consumer fe
Source: 机器之心
Read on 88z.aiOpenAI has announced the availability of Locked Mode for all individual ChatGPT users, disabling high-risk online features while allowing access to cached content and image processing.
Source: ChatGPT
Read on 88z.aiDaxiao Robotics, in collaboration with The Chinese University of Hong Kong and Shenzhen He Tao College, has launched Kairos-HomeWorld, the world's first whole-house 3D interactive world model. Built on a four-stage hierarchical generation architecture, the model generates coherent, physically plausible, and functionally complete 3D home environments from text prompts with a single click. Each scene averages over 15 interactive objects, enabling robots to undergo simulation training for complex h
Source: 商汤科技SenseTime
Read on 88z.aiGoogle and SpaceX have reached a cloud computing partnership. From October 2026 to June 2029, Google will pay SpaceX $920 million monthly to lease computing power from approximately 110,000 Nvidia GPUs and CPUs for AI training and inference. The agreement aims to ease Google's compute supply constraints while creating a significant revenue stream for SpaceX's AI business, adding new narrative value to its IPO.
Source: IT之家
Read on 88z.aiNext-generation AI ultrasound BCI company Huachao Shenkong has completed a hundred-million-yuan angel series financing round, backed by Matrix Partners China, Delian Capital, and Daoyuan Capital. The funds will support the verification and industrialization of its low-intensity focused ultrasound non-invasive brain-computer interface technology, advancing clinical translation for indications such as pain and addiction.
Source: 36氪
Read on 88z.aiOpenAI has launched a new ChatGPT memory system based on Dreaming technology, which automatically organizes and updates user background information for more personalized responses. The new feature supports viewing and editing memory summaries and is initially available to Plus and Pro users in the US.
Source: APPSO
Read on 88z.aiTencent has launched WorkBuddy Enterprise Edition, offering digital employees and human-machine collaboration modes. It supports various deployment options and integrates with Tencent's office ecosystem, aiming to upgrade efficiency from individual to organizational levels by accumulating AI assets.
Source: 腾讯WorkBuddy
Read on 88z.aiTongyi Lab has released PawBench, a general agent benchmark that jointly evaluates base models and execution harnesses for the first time. The benchmark includes numerous real-world tasks and test units across various model and harness combinations. Results show significant performance gaps between different harnesses, with score differences for the same model reaching up to 11.5 points, highlighting the critical impact of harness selection on agent performance.
Source: 通义实验室
Read on 88z.aiGoogle's Magenta team releases Magenta RealTime 2, offering two versions optimized for Apple Silicon. It significantly reduces latency and supports various control methods, accompanied by new tools for low-latency improvisation on Mac.
Source: IT之家
Read on 88z.aiAI music unicorn Suno has completed a $400 million funding round, doubling its valuation to $5.4 billion. The company has surpassed 2 million subscribers, with annual sales projected to reach $300 million, and plans to expand its workforce by 70% by year-end. Additionally, Suno has reached a settlement and licensing agreement with Warner Music.
Source: Z Potentials
Read on 88z.aiMoonshot AI has released the beta version of Kimi Work, a general-purpose local agent designed for knowledge workers and integrated into the latest Mac and Windows clients. Built on the Kimi Code core, it supports skill installation, scheduled tasks, and browser automation, while coordinating up to 300 sub-agents. It is currently in closed beta for finance, research, and office scenarios.
Source: 月之暗面Kimi
Read on 88z.aiJD.com launches JoyAI-Echo, a long audio-video generation framework. Through four technical innovations including cross-modal memory banks, it addresses challenges in character consistency, voice stability, and generation speed, with benchmark results showing comprehensive leadership.
Source: 京东黑板报
Read on 88z.aiNous Research has released a preview of the official Hermes Desktop client, supporting macOS, Windows, and Linux. The tool integrates core features such as persistent memory, task scheduling, and multimodal reasoning, allowing users to run it directly on their local desktop without terminal commands, marking Hermes' evolution into a more user-friendly desktop AI agent.
Source: Desktop
Read on 88z.aiOpenAI announced three updates for its coding agent Codex: pre-set job skill packs for no-code AI employee deployment, precise editing of generated content, and one-click conversion of documents into interactive websites. Codex has reached 5 million weekly active users, an eightfold increase since the start of the year, with plans to integrate into ChatGPT and support one-click deployment on AWS.
Source: 智东西
Read on 88z.aiCognition announced that Windsurf has been upgraded to Devin Desktop, transforming into an Agent command center. The new version uses Kanban to unify local and cloud Agents, adds Spaces for context sharing, and supports the ACP protocol for third-party Agent integration. It retains full IDE capabilities, allowing existing users to upgrade automatically via OTA for zero-cost migration.
Source: Devin
Read on 88z.aiMicrosoft has released seven self-developed MAI models trained from scratch, covering reasoning, image, voice, and coding domains, all using unified data standards and evaluation frameworks.
Source: Microsoft.ai
Read on 88z.aiQianwen App has announced full openness to third-party Agents and Skills, allowing enterprises to operate brand-specific intelligent agents. Initial partners such as KFC and Luckin Coffee are currently testing, with services rolling out gradually. These agents feature memory and proactive planning capabilities to offer personalized services like itinerary reminders and repurchase recommendations.
Source: 千问APP
Read on 88z.aiByteDance has officially open-sourced the Bernini framework, designed to unify video generation and editing workflows. The framework innovatively employs a multimodal large model as its planning core, which precisely understands user intent and semantic goals to guide diffusion models in producing high-quality visual renders, thereby significantly enhancing the controllability of AI video editing.
Source: 量子位
Read on 88z.aiAlibaba Tongyi has launched Qwen3.7-Plus, a multimodal agent foundation model that unifies visual and language capabilities to support scene perception, GUI operation, and code generation. The model ranks fifth globally and first in China on the Vision Arena leaderboard, demonstrating the ability to handle complex workflows such as full-link app development.
Source: 千问大模型
Read on 88z.aiThe Peking University team launched ChatExcel Max 2.0, removing file limits while preserving original formats. It supports conversational data processing and multi-format report generation, enabling automated analysis and data security for finance, logistics, and other sectors.
Source: 元空AI
Read on 88z.aiAlibaba DAMO Academy has launched Yugou, an AI-native development platform centered on Vibe Coding that enables users to generate complete digital products via natural language. The platform covers the entire workflow from requirement understanding to release and features built-in standardized resource interfaces for zero-code access.
Source: 语构
Read on 88z.aiGoogle parent company Alphabet has announced an $80 billion equity financing round, primarily directed toward AI infrastructure and computing power. The funding comes from public offerings, at-the-market issuances, and a private investment by Berkshire Hathaway. This move aims to support the company's growing capital expenditure plans and strong cloud business demand.
Source: IT之家
Read on 88z.aiCiyuan Wuxian has completed a tens-of-millions angel+ round of financing, co-led by Huakong Fund and Shuimu Ventures, with Xiajin Innovation participating. The founding team combines expertise from ByteDance's AI transformation projects and Tsinghua University's Yao Class. Focusing on enterprise-grade Agent infrastructure, its InfCode and InfOne products already serve leading financial and software companies.
Source: Founder Park
Read on 88z.aiCoze 3.0 has officially launched, covering iOS, Android, Mac, Windows, and web platforms. The new version focuses on AI team collaboration, supporting multi-Agent collaboration and project space management. It allows access to local Agents like Claude Code, provides long-term cloud running capabilities, and includes various professional templates.
Source: 扣子Coze
Read on 88z.aiMiniMax has released the open-source MiniMax M3, the first domestic model combining advanced coding, million-token context, and native multimodal capabilities. Utilizing a new MSA sparse attention architecture, it achieves strong performance in international benchmarks, surpassing some leading overseas models.
Source: MiniMax 稀宇科技
Read on 88z.aiThe HaiSnap app has officially launched as a mobile AI creation agent platform. Users can initiate tasks via voice, text, or file uploads, with AI automatically generating publishable outputs such as web apps, documents, PPTs, and images. Core features include an agent workspace, results library, task playback, and skill invocation.
Source: 响指HaiSnap
Read on 88z.aiGeneral AI company VAST has completed its A+ and A++ funding rounds totaling nearly $200 million and launched Project Eden, a world model that decouples state inference from visual rendering.
Source: VAST AI
Read on 88z.aiXiaomi's large model application team has released the open-source model ControlFoley, which uniformly supports three video dubbing tasks: text-guided, text-controlled, and reference audio-controlled. By leveraging visual encoding and modal robust training, the model addresses issues of weak control and synchronization difficulties, achieving state-of-the-art performance among open-source solutions on multiple benchmarks.
Source: 小米技术
Read on 88z.aiTongyi Lab releases Qwen-VLA, built on Qwen3.5-4B and a DiT action decoder, integrating manipulation, navigation, and trajectory prediction to support 11 robot platforms.
Source: 通义实验室
Read on 88z.aiAnthropic has completed a $65 billion Series H funding round, reaching a post-money valuation of $965 billion. The round was led by Altimeter Capital and Sequoia Capital, with participation from cloud providers like Amazon, and is expected to be the final private round before an IPO.
Source: IT之家
Read on 88z.aiAnthropic has released its flagship model, Claude Opus 4.8, featuring significant enhancements in terminal engineering and knowledge work, with some metrics surpassing the unreleased Mythos. The core breakthrough lies in improved honesty, reducing code defect false negatives to one-quarter of the previous generation and overconfidence probability to one-tenth, earning high praise from the CEOs of Cursor and Devin.
Source: 量子位
Read on 88z.aiTongyi Lab has released Qwen-Image-Bench, a text-to-image evaluation benchmark featuring 56 fine-grained checkpoints and the open-source Q-Judger assessment model. Developed with professional artists, it covers real-world scenarios like world knowledge and creative reasoning to bridge the gap between basic generation and professional creation.
Source: 通义实验室
Read on 88z.aiAstronClaw on the iFlytek Open Platform has launched Agent Team mode, enabling automatic formation of multi-specialist agent teams via single-sentence commands. It offers pipeline, map-reduce, and supervisor collaboration mechanisms to parallelize complex task breakdown and delivery. With a visual dashboard and support for models like GLM 5.1 and Kimi 2.6, it drastically reduces execution time for long-cycle tasks such as product planning.
Source: 讯飞开放平台
Read on 88z.aiAlibaba Cloud has officially open-sourced the Bailian CLI, enabling developers to connect AI Agents to over 100 models and applications on the Bailian platform with a single command. It supports full capabilities including knowledge bases, memory, and web search. Designed for Agent scenarios, it features multimodal calls and native compatibility with mainstream frameworks like Claude Code, Qoder, and OpenClaw.
Source: 阿里云
Read on 88z.aiModelBest, together with Tsinghua THUNLP and OpenBMB, has open-sourced the agent operating system PilotDeck. Replacing traditional chat interfaces with workspaces, it addresses key pain points such as opaque agent memory, uncontrollable costs, and passive task execution.
Source: 面壁智能
Read on 88z.aiOn its third anniversary, 01.AI founder Li Kaifu announced a 'second startup' phase, shifting focus from AGI ideals to industrial AI implementation. Benchmarking against Palantir, the company targets high-margin, high-retention B2B services, aiming to become China's first profitable AI 2.0 firm by a quarter next year. New measures include a partnership system, 20 million stock options, CEO incentives, and the DRI framework.
Source: 零一万物 01AI
Read on 88z.aiTencent has launched the AI game creation platform 'Code Craft', which supports various game genres, offers a vast library of free assets and AIGC capabilities, and is recruiting initial beta testers through an offline training camp co-launched with Shenzhen Qianhai.
Source: 代号Craft
Read on 88z.aiRunway has launched the Runway MCP Server, integrating its AI video and image generation capabilities with major agent-compatible platforms. Users can directly invoke models such as Gen-4.5, Seedance 2.0, and GPT Image 2 within conversations to create marketing videos, product assets, and visual content without switching contexts.
Source: Runway
Read on 88z.aiX Square Robot has released the open-source embodied large model Wall-OSS-0.5, which supports direct deployment after pre-training. The model performed well in zero-shot tests across 17 real-world robot tasks, with scores exceeding 80 in four of them.
Source: 机器之心
Read on 88z.aiMicrosoft Research releases MAI-Image-2.5, rising to third place in text-to-image rankings and breaking the monopoly of tech giants. The model significantly enhances text rendering and visual details, suitable for commercial design scenarios.
Source: IT之家
Read on 88z.aiAlipay has released a full-stack AI-native payment system, introducing the world's first Token Pay service and an AI wallet. Together with its existing AI Pay and AI Collect features, this forms a complete infrastructure for the agent economy, covering authorization, payment, settlement, and management. AI Pay has processed 300 million transactions with minimal loss rates, while AI Collect enables agents to automatically purchase services and monetize personal expertise.
Source: 智东西
Read on 88z.aiXiaomi's MiMo-V2.5 series API announces a permanent price cut, with reductions up to 99% and Token Plan usage increased by 5 to 8 times. The trillion-token incentive plan ended early, distributing 100 trillion tokens to over 548,000 applicants. Existing user quotas are reset, and the V2.5-TTS series remains free for a limited time.
Source: Xiaomi MiMo
Read on 88z.aiKuaishou has officially open-sourced its self-developed multimodal large model, Keye-VL-2.0-30B-A3B. It introduces the DSA sparse attention mechanism to multimodal understanding for the first time, supporting a 256K ultra-long context window and enabling temporal causal reasoning for long videos. The model surpasses Gemini 2.5 Pro and Gemini 3 Flash on certain metrics in the TimeLens benchmark and unlocks Agent collaboration capabilities for complex tasks like coding, tool use, and search.
Source: 量子位
Read on 88z.aiKunlun Wanwei's Tiangong AI releases SkyClaw-v1.0 and its lite version, supporting million-token context, optimized for complex tool use and code generation, outperforming mainstream open-source models at a lower price.
Source: 昆仑万维集团
Read on 88z.aiXiaomi EV has released a new framework that deeply couples 3D reconstruction with video generation. The reconstruction module uses sparse 3D anchors, processing 10-second videos in just 10 seconds. The generation module achieves efficient output of 0.19 seconds per frame with 4-step denoising via two-stage training and ODE distillation, supporting videos up to one minute long.
Source: 小米技术
Read on 88z.aiAlibaba Tongyi Lab releases AgentScope 2.0, featuring systematic upgrades for production stability. The new version introduces model fault tolerance, real-time message visibility, permission-based interception of high-risk operations, and context optimization to enhance multi-agent reliability in real-world scenarios.
Source: 通义实验室
Read on 88z.aiModelBest, together with Tsinghua University and the OpenBMB open-source community, has released MiniCPM5-1B. The model scored 17.9 on the AA-Index leaderboard, surpassing all models under 2B parameters and some larger ones, making it the strongest open-source foundation model within the 2B parameter range globally. After INT4 quantization, its size is only 0.5GB, enabling smooth operation on terminals such as mobile phones and browsers.
Source: 面壁智能
Read on 88z.aiModelBest, together with Tsinghua University and the OpenBMB open-source community, has released BitCPM-CANN, China's first 1.58-bit ternary large model trained end-to-end entirely on Huawei's Ascend domestic computing platform. Available in four sizes ranging from 0.5B to 8B, the model uses quantization-aware training to free up approximately six times the video memory during inference while retaining 90% to 97.2% of its capabilities. This release aims to provide critical infrastructure for edg
Source: 面壁智能
Read on 88z.aiSenseTime has officially launched the Raccoon Desktop Agent, positioning it as an AI office partner that evolves from an advisor to an executor. The product features six core capabilities, including direct reading of over twenty local file formats, browser automation, global shortcut activation, office collaboration integration, local memory learning, and automated scheduled tasks, aiming to enhance office efficiency.
Source: 商汤科技SenseTime
Read on 88z.aiTencent ima copilot has announced full open access, eliminating wait times. The tool features memory and full-scenario awareness, supporting knowledge base access and integration with major model APIs. The Knowledge Square has upgraded to a capability platform, adding Skill publishing functions and launching official skills like WeChat Read.
Source: ima.copilot
Read on 88z.aiDeepSeek has announced a permanent price reduction for the DeepSeek-V4-Pro model API. After the limited-time offer ends on May 31, the new rate will be fixed at one-quarter of the original price, significantly lowering long-term costs for developers.
Source: IT之家
Read on 88z.aiZhipu Qingyan has officially launched the AgentMore Skills Square, integrating three major modules: official recommendations, a vast library of free skills, and an open-source community. It supports global search and one-click installation with zero tokens. Users can quickly equip agents with exclusive skill packs such as WeChat Reading and Meituan coupons, enabling instant upgrades from general chat to industry expertise, with flexible uninstallation and management options.
Source: 智谱清言
Read on 88z.aiZhipu Open Platform has released the GLM-5.1 High-Speed API, achieving an output speed of 400 tokens/s and setting a new global benchmark for large model API speed. Jointly launched by Zhipu and the TileRT team, it combines flagship capabilities with ultra-low latency. It is suitable for high-demand scenarios like AI coding and real-time interaction, and is now available to select enterprise clients.
Source: 智谱开放平台
Read on 88z.aiMeituan's LongCat team releases LongCat-Video-Avatar 1.5, significantly improving lip sync and video stability by upgrading the audio encoder, optimizing the data system, and introducing GRPO alignment. The model uses DMD distillation for 8-step generation, boosting efficiency by approximately 15 times, and advancing digital human technology from open-source SOTA to commercial-grade applications.
Source: 龙猫LongCat
Read on 88z.aiNetEase Youdao's AI team has open-sourced Confucius4, a multimodal reasoning model based on the Qwen3.5-27B architecture. Designed for advanced mathematical reasoning, it supports mixed text-image inputs and is compatible with the Qwen ecosystem.
Source: Confucius4
Read on 88z.aiAlibaba has launched QoderWake, a production-grade AI digital employee platform, initiating a global public beta with an initial Mac release. The platform allows for rapid configuration in five minutes and includes six built-in roles such as backend and frontend development. Featuring long-term memory and evolutionary capabilities, digital employees can accept tasks via IM or Webhooks and automatically generate delivery summaries. Security is ensured by a four-dimensional permission center, whil
Source: Qoder
Read on 88z.aiZhipu, Yuxun Network, and Tsinghua University launched ZCube, an architecture that eliminates the Spine layer to address PD separation inference congestion. Tests show it significantly reduces hardware costs while boosting GPU throughput and response speed, providing an efficient foundation for ultra-large-scale inference clusters.
Source: 智谱
Read on 88z.aiTicket sales have opened for the 2026 Ali D20 Global Design Dean Summit, with Duiyou participating once again. Centered on the theme of Design Rebirth, the main forum addresses large models, Agents, AIGC, and spatial intelligence. Sub-forums cover cutting-edge areas such as Alibaba Cloud Vibe Designing, Taotian AI Design Agent, and Qwen AI hardware UX, while introducing AI design workshops and exhibitions for the first time.
Source: 堆友
Read on 88z.aiVolcengine Ark's Coding Plan and Agent Plan now support the DeepSeek V4 model series. Users can access both DeepSeek V4 Pro and V4 Flash within these packages. The Coding Plan is tailored for developer coding scenarios, while the Agent Plan integrates multimodal models like DeepSeek V4 and Doubao Seed, along with Harness capabilities such as web search. Subscription-based pricing significantly reduces invocation costs.
Source: 火山引擎
Read on 88z.aiTencent Hunyuan has released the open-source Hy-MT2 translation model in three sizes, supporting multilingual and dialect translation with top-tier performance in certain benchmarks. The Hy Translation mini program was also launched, featuring voice input, custom styles, and offline capabilities, with mobile apps coming soon.
Source: 腾讯混元
Read on 88z.aiTencent's OS-level AI assistant Marvis has officially launched, offering multi-platform downloads without an invitation code. Developed by the Yingyongbao team, it features always-on AI agents capable of reading local files, controlling PC software and mobile apps, and providing system-level file MCP services.
Source: IT之家
Read on 88z.aiAt I/O 2026, Google launched Gemini 3.5 Flash and Omni Flash models, upgraded search, introduced AI glasses and development platforms, marking the entry into an agent-driven Gemini era.
Source: 智东西
Read on 88z.aiAlibaba Tongyi Qwen has released the flagship model Qwen3.7-Max, positioned as a versatile foundation for the agent era. The model excels in programming, office automation, and long-cycle autonomous execution, with benchmarks showing it leads peers in multiple agent evaluations.
Source: 千问大模型
Read on 88z.aiAlibaba Tongyi Qwen releases the real-time simultaneous interpretation model Qwen3.5-LiveTranslate, supporting multi-language audio input and output with low-latency translation, voice cloning, and industry terminology optimization.
Source: 通义实验室
Read on 88z.aiAt the 2026 Alibaba Cloud Summit, Alibaba Cloud unveiled Qwen Cloud, a new platform designed for AI Agents. It integrates APIs from over 150 mainstream models, including Qwen, DeepSeek, and Kimi. The platform transforms model services into Skills and CLI tools, allowing Agents to automatically handle model selection, invocation, and task execution with a single command, eliminating the need for manual integration code.
Source: 阿里云开发者
Read on 88z.aiHiDream has open-sourced its flagship image generation model, HiDream-O1-Image-Pro. Built on the UiT native multimodal architecture, it maps pixels, text, and task conditions into a shared token space, eliminating traditional VAEs and separate encoders. The model achieves state-of-the-art results on benchmarks like GenEval and DPG, ranking eighth in the Artificial Analysis Arena and surpassing open-source models such as FLUX.2[Dev].
Source: HiDream智象未来
Read on 88z.aiWeChat Read has officially launched WeChat Read Skill. After binding an API key, AI can access personal reading data to perform six functions including checking bookshelves and searching for books. The skill supports deployment to assistants like WorkBuddy, enabling natural language queries and batch export of notes to Obsidian and Notion, thereby activating the value of personal reading data.
Source: APPSO
Read on 88z.aiLucius has completed a multi-million dollar angel financing round led by Mingshi Venture Capital, with Blue Lion Capital serving as the exclusive financial advisor. The company focuses on addressing the pain point of missing organizational memory in enterprises. Its AI Teammate records questions rather than just storing answers and has served over 30 clients. A typical case shows that after Dubbing AI integrated the solution, its community self-resolution rate jumped from 29% to 88%.
Source: Z Potentials
Read on 88z.aiElon Musk invited users on X to test Cursor's latest model, Composer 2.5, which was partially trained using Colossus 2. The official statement indicates it builds on Moonshot AI's Kimi K2.5, focusing on improving stability for long tasks and adherence to complex instructions, while enhancing performance through directed reinforcement learning and expanded synthetic training.
Source: IT之家
Read on 88z.aiTencent's App Bao team has launched internal testing for Tencent Marvis, an OS-level AI assistant. It supports the local deployment of six collaborating agents and uses natural language commands for file management, system settings, and cross-modal retrieval, balancing cloud efficiency with local privacy.
Source: 腾讯科技
Read on 88z.aiThe open-source AI agent OpenHuman has continuously topped the GitHub Trending charts, garnering over 9,000 Stars. Its popularity is driven by features such as zero-config setup, one-click connectivity to over 100 services, and significant reductions in token consumption.
Source: 量子位
Read on 88z.aixAI has released Grok Build, its first terminal-native AI coding agent, available to SuperGrok Heavy subscribers. The tool supports parallel sub-agents, integrates with existing workflows and MCP, and features planning modes along with in-terminal image and video generation. Concurrently, Elon Musk has merged xAI into SpaceX and updated its logo.
Source: 量子位
Read on 88z.aiLibTV Team Edition has officially launched, positioned as a professional AI video collaboration workspace. It offers six key features: dedicated project spaces, real-time multi-user canvas collaboration, a unified asset library, tiered permission management, a shared credit pool, and automatic member asset handover. The service targets AI video studios, short drama producers, brand marketing departments, and enterprise content teams.
Source: 哩布哩布AI
Read on 88z.aiTencent has launched 'Tusi', an exploratory vibe coding product positioned as an app generation and inspiration co-creation platform. Users without coding skills can use natural language descriptions to let AI handle the entire process from function breakdown to app packaging. The platform supports app creation, social sharing, replication in the inspiration square, and AI app search, aiming to build a complete loop of idea-to-app distribution and co-creation.
Source: 吐司
Read on 88z.aiTencent's AI design tool Ardot has launched its full public beta, targeting product, design, and R&D teams. It supports UI generation from text or images, local edits, and batch icon creation. Via the MCP protocol, it converts designs to code, compatible with major IDEs. The tool features collaboration, permission management, and deep Figma compatibility with enhanced performance.
Source: 腾讯设计Ardot
Read on 88z.aiOpenAI announces that Codex has officially launched on the ChatGPT mobile app, with previews available simultaneously on iOS and Android, accessible even to free users. The feature supports code approval, thread viewing, and project initiation on mobile devices, with real-time status synchronization. Additionally, tools such as remote SSH connections and programmatic access tokens are introduced to enhance team automation and security management capabilities.
Source: 机器之心
Read on 88z.aiMoonshot AI has released the Kimi WebBridge browser extension, enabling local AI agents like Kimi Code and Claude Code to directly operate browsers. By carrying user login states and cookies, the tool automates tasks such as clicking, form filling, and cross-site information integration, requiring only the installation of the extension and a local daemon.
Source: 月之暗面 Kimi
Read on 88z.aiAlibaba Cloud has released Qoder 1.0, upgrading the product from an AI integrated development environment to an autonomous agent development workbench. The new version introduces a Quest independent window for cross-project multi-task parallel processing, integrates memory, Repo Wiki, and knowledge cards to build a team-level knowledge engine, and supports Experts groups and custom Agent teams. The version is currently available for download on Windows, macOS, and Linux platforms.
Source: 阿里云
Read on 88z.aiAlibaba Cloud has released Wanxiaozhi 2.0, an enterprise-grade AI website building platform, featuring an upgraded brand logo and virtual avatar. The new version covers the entire website creation process, supporting industry templates, one-click setup from reference sites, and professional requirement document generation. It natively integrates domain registration, ICP filing, SSL certificates, and one-click deployment for compliant launch. Equipped with a visual management backend and intellig
Source: 阿里云万网
Read on 88z.aiModao AI introduces a new feature that generates high-fidelity apps from natural language descriptions. The system proactively clarifies requirements, matches mainstream design systems, and generates frontend code for rapid delivery from concept to development.
Source: 墨刀
Read on 88z.aiXiaomi's Embodied AI team has open-sourced Xiaomi OneVL, a one-step latent space language-vision reasoning framework. The model is the first in the industry to unify VLA, world models, and latent space reasoning, combining strong reasoning capabilities with high-speed response. Its accuracy surpasses explicit CoT, while its speed aligns with latent space CoT solutions. It achieves SOTA on multiple benchmarks including ROADWork, Impromptu, and Alpamayo-R1, offering dual-dimensional interpretabili
Source: IT之家
Read on 88z.aiRunway has officially launched Runway Agent, a conversational AI video creation agent. By simply describing their needs, users can have the agent handle the entire workflow from concept to final product in a single turn of dialogue, automatically generating high-resolution videos with multiple scenes, voiceovers, and soundtracks.
Source: Runway
Read on 88z.aiByteDance's Doubao Input Method has officially launched its macOS version, enabling Apple computer users to input text via voice. Built on the same speech model as the Doubao app, it features real-time transcription without time limits, supporting various dialects, mixed Chinese-English speech, and intelligent error correction.
Source: 豆包
Read on 88z.aiAt the Android Show, Google launched the first Googlebook laptop designed specifically for Gemini, featuring an intelligent cursor and cross-device synergy. It also released Gemini Intelligence for Android, enabling cross-app task automation, Chrome plugin integration, speech-to-text, and custom desktop widgets. Additionally, Google partnered with Apple to add end-to-end encrypted RCS and previewed smart glasses for the upcoming I/O conference.
Source: 智东西
Read on 88z.aiMiniMax has released Mavis, an upgraded version of MiniMax Agent, featuring Agent Teams for multi-role parallel collaboration. It addresses single-agent stagnation and drift via an Owner, Worker, and Verifier architecture. Mavis also merges TokenPlan and Agent Plan, allowing a single subscription to cover the full CLI, API, and Agent workflow, supporting M2.7 and audio-video models with flexible shared Credits.
Source: MiniMax 稀宇科技
Read on 88z.aiFormer Alibaba Qwen technical lead Lin Junyang has officially launched a startup focusing on world models and embodied brains. The new company is valued at around $2 billion, with core team members from ByteDance, Tencent, and overseas backgrounds, and is currently seeking funding from investors such as Sequoia China.
Source: 智能涌现
Read on 88z.aiAnthropic has launched a research preview of Agent View for Claude Code, enabling the scheduling and monitoring of multiple parallel AI sessions within a single terminal. This feature allows users to check task status in real time, switch sessions seamlessly without losing progress, and upgrades the tool into a multi-task command center available to all paid plans.
Source: Claude
Read on 88z.aiOpenAI announces the formation of a new company and the launch of the Daybreak cybersecurity tool, aiming to deepen enterprise AI deployment and automated defense capabilities.
Source: 机器之心
Read on 88z.aiSenseTime's Office Raccoon Desktop 2.0 has been released, transforming into an execution assistant with six new capabilities including local file reading and browser control. The Mac version has opened beta testing, while the Windows version is expected to launch in mid-to-late May.
Source: 商汤小浣熊
Read on 88z.aiOpenBMB releases MiniCPM-V 4.6, an edge multimodal large model with 1.3B parameters and ~1.6GB size, supporting pure local offline deployment on iOS, Android, and HarmonyOS NEXT.
Read on 88z.aiCodeFlying announces an upgrade to a one-stop platform for AI-assisted business, supporting Chinese prompts to generate mini-programs, apps, and websites, with dedicated AI staff for operations and auto-generated admin panels.
Source: 码上飞-CodeFlying
Read on 88z.aiQwen App has fully integrated with Taobao. After updating to version 6.9.1, users can select, compare, and purchase products directly within the app. Leveraging AI capabilities, Qwen supports precise filtering under complex conditions, intelligent understanding of vague requirements, and scenario-based product recommendations. It can also identify overpriced items to advise against purchase, creating a full-link AI shopping experience from search to checkout.
Source: 千问APP
Read on 88z.aiEmbodied AI infrastructure company Uncharted Dynamics has completed a multi-million dollar seed funding round led by Xianfeng Changqing. The company is dedicated to developing high-precision multibody dynamics solvers, aiming to address physical distortion in simulations and provide a reliable physical foundation for embodied AI.
Source: Z Finance
Read on 88z.aiAt the 2026 Mobile Cloud Conference in Suzhou, China Mobile unveiled innovations including the desktop AI office agent MobileClaw, the MoMA large model aggregation platform, and a mobile AI cloud PC supporting phone interactions. The company also introduced AI-eSIM cards, a trillion-token service experience package, and plans for gigawatt-scale AI data parks.
Source: 中国移动
Read on 88z.aiVolcengine launches the Agent Plan package, integrating multimodal models with Harness tools, supporting mainstream models like Doubao-Seed and GLM-5.1, and platforms such as Claude Code.
Source: 火山引擎
Read on 88z.aiGoogle Cloud Agents has released the open-source command-line tool agents-cli, designed to streamline the entire lifecycle of enterprise-grade AI Agents from development to deployment. The tool integrates various mainstream programming agents, includes seven skill packages, and supports local evaluation, automated deployment, and human takeover capabilities.
Source: agents-cli
Read on 88z.aiTencent QClaw has launched the 'File Space' feature, enabling interoperability with Tencent Docs and the ima knowledge base. Users can directly access local files, Tencent Docs, and ima resources within QClaw for AI processing. Generated content can be saved as a Tencent Doc with a collaboration link or stored in the ima knowledge base, creating a closed loop from data retrieval and AI processing to team collaboration.
Source: 腾讯云
Read on 88z.aiStepFun has released the new generation real-time voice large model StepAudio 2.5 Realtime, featuring three core advantages: top-tier paralinguistic capabilities, millions of customizable personas, and leading conversational EQ and IQ.
Source: 阶跃星辰
Read on 88z.aiBaidu has officially released ERNIE 5.1, significantly reducing pre-training costs by compressing parameter scale. The model performs excellently in various benchmarks, ranking fourth globally and first domestically, while demonstrating strong competitiveness in agent capabilities and math competition scores.
Source: 百度文心
Read on 88z.aiAnt Bailin releases the trillion-parameter model Ring-2.6-1T, significantly enhancing reasoning and multimodal capabilities while reducing inference costs. The API is now open for various task scenarios.
Source: 百灵大模型心
Read on 88z.aiTongyi Lingma has launched the RepoWiki feature, which automatically generates structured documentation covering module relationships, interface definitions, and implementation logic. This addresses the issue of fragmented code understanding in AI programming tools. The feature supports full generation, incremental updates, and Git directory synchronization, allowing Wiki files to be shared with teams via code commits. Currently in Beta, it is available for free to Enterprise Standard and Exclus
Source: 灵码Lingma
Read on 88z.aiOpenAI releases a limited preview of GPT-5.5-Cyber, optimized for security tasks via the TAC framework and available to vetted defenders.
Source: OpenAI
Read on 88z.aiOpenAI has released three real-time voice models that enhance reasoning and tool use, offer low-cost multilingual translation, and provide low-latency transcription, all delivered via the Realtime API with end-to-end emotion preservation.
Source: openAI
Read on 88z.aiSenseTime has released the new lightweight multimodal agent model, SenseNova 6.7 Flash-Lite. Built on a native multimodal architecture, it integrates perception, reasoning, and action, reducing token consumption by 60% compared to pure text agents. It achieves state-of-the-art performance in its class across multiple authoritative benchmarks. Additionally, SenseTime launched a limited-time free SenseNova Token Plan, allowing 1,500 calls every five hours.
Source: 商汤科技SenseTime
Read on 88z.aiMeituan has launched the public beta of 'Miyou,' an AI-native community designed for large models and Agents. Featuring a pioneering 'shrimp-raising' mode, it allows users to raise AI Agent shrimps that help find income opportunities and connect with peers. The platform currently hosts over 3,000 Agents and aggregates more than 40,000 Skills.
Source: 觅游
Read on 88z.ai360 Group's 360 AI Office Platform has launched JJT, an AI PPT tool that supports various creation modes and format compatibility, featuring intelligent editing and multi-format export capabilities.
Source: JJT
Read on 88z.aiGongji Technology announced the completion of a Pre-A financing round of nearly 100 million yuan, jointly led by Chunhua Ventures and Bi'an Shidai. The company integrates global idle and heterogeneous computing resources through its proprietary power-grid-style computing scheduling network. By using AI algorithms to dynamically match supply and demand, it allows computing power to be used on demand and billed by volume, similar to electricity. The platform currently serves over 5,000 teams and e
Source: 共绩算力suanli.cn
Read on 88z.aiOpenAI has officially launched the Codex Chrome extension, enabling Codex to directly operate within a user's logged-in Chrome browser. The plugin supports reading tabs, page content, and login information, while performing interactions like navigation, clicking, typing, and screenshots. Users retain full control, with the ability to stop tasks, manage permissions, and confirm interactions with new websites.
Source: OpenAI
Read on 88z.aiAt the Code with Claude developer conference, Anthropic announced several updates, including increased usage limits for Claude Code, relaxed API rate limits, enhanced managed agent capabilities, and a compute partnership with SpaceX.
Source: 智东西
Read on 88z.aiMoonshot AI (Kimi) has completed a Series D financing round of approximately $2 billion, pushing its post-money valuation above $20 billion and setting a record for the largest single funding deal in China's large model sector. Led by Meituan Dragonball with participation from China Mobile and others, the company's cumulative funding exceeds $4.4 billion, ranking first among domestic startups. In April, its annual recurring revenue surpassed $200 million, alongside the release of the open-source
Source: 智东西
Read on 88z.aiAccording to the Financial Times, China's National Integrated Circuit Industry Investment Fund (Big Fund) is negotiating to lead an investment in DeepSeek, with the current valuation nearing $45 billion, doubling from two weeks ago. Founder Liang Wenfeng may personally co-invest, while Tencent and Alibaba are also discussing participation. This marks the Big Fund's first investment in an AI large model company.
Source: 智东西
Read on 88z.aiByteDance has open-sourced Mamoda2.5, the world's first 25B-class unified multimodal model. Built on the Qwen3-VL-8B architecture, it achieves state-of-the-art performance in text-to-image, video generation, and editing, with inference speeds significantly surpassing competitors and performance nearing top closed-source models.
Source: Mamoda2.5
Read on 88z.aiInfinigence AI announced a new funding round exceeding 700 million yuan, jointly led by Hangzhou High-Tech Investment Group and Huiyuan Capital, maintaining its lead in financing scale among China's AI-native infrastructure companies. The company also released the 'AI Productivity Formula,' positioning itself as a hub for the token economy, with daily token calls on its Agentic MaaS platform growing more than 20-fold since late last year.
Source: 无问芯穹
Read on 88z.aiOpenAI, together with AMD, Broadcom, Intel, Microsoft, and NVIDIA, has launched the Multi-path Reliable Connection (MRC) protocol, open-sourced to the industry via OCP. Built on RoCE extensions, MRC uses multi-plane network design and adaptive packet spraying to distribute data across hundreds of parallel paths. This approach prevents congestion and hotspots, reducing fault recovery time from seconds to microseconds.
Source: 量子位
Read on 88z.aiPixelBloom has completed its Series C funding, co-led by Guoke Investment and SenseTime Guoxiang Capital. The company has transformed into an AI office solution Agent, launching the world's first marketing solution Agent, 'Xiao Fang Tongxue,' shifting from content generation to result delivery. Its consumer product AiPPT has reached 30 million users, while its B2B services are integrated into hardware ecosystems like Lenovo and Huawei, as well as enterprises such as CITIC Securities and Amway Ch
Source: 像素绽放PixelBloom
Read on 88z.aiOpenAI has released GPT-5.5 Instant, replacing GPT-5.3 Instant as the default model for ChatGPT. The new model significantly reduces hallucinations in medical, legal, and financial fields, enhances mathematical capabilities, and provides more concise and natural responses. Plus and Pro users can leverage historical conversations and files for personalized answers, while all consumer versions now include a feature to manage memory sources.
Source: APPSO
Read on 88z.aiDeepSeek-TUI, launched by US developer Hunter Bown, has gained popularity on GitHub with 2.3k stars. Built in Rust under the MIT license, this terminal coding agent is optimized for DeepSeek V4 and supports file operations, shell execution, Git management, and MCP integration.
Source: 量子位
Read on 88z.aiByteDance's AI assistant Doubao has launched a test for paid subscription services with three pricing tiers while maintaining a free basic version. The app currently boasts over 140 million daily active users, leading the domestic AI assistant market.
Source: 量子位
Read on 88z.aiOpenAI has launched the ChatGPT Advertiser Platform, enabling ad buying via agencies or self-serve tools, with new CPC bidding and measurement capabilities.
Source: OpenAI
Read on 88z.aiiFlytek Zhiwen has introduced the Vision Agent mode, transforming AI's role in PPT creation from passive generation to active collaboration through intent insight, outline construction, and a five-step design engine. This mode addresses issues like rigid templates and mismatched images by supporting online research, autonomous page structure planning, and editable content generation, resulting in professionally formatted presentations.
Source: 讯飞开放平台
Read on 88z.aixAI has released Grok 4.3, positioned as a pragmatic transitional version. The model features significantly reduced API pricing, faster output speeds, and support for million-token long contexts, with enhanced performance in document generation and office assistance tasks. While showing marked improvement in agent tasks, it still lags behind top competitors in complex reasoning and fact-checking stability, with an increased hallucination rate.
Source: 新智元
Read on 88z.aiOpenAI's coding tool Codex introduces a pet mode, allowing users to keep a cyber pet on their desktop. Activated via the /pet command, it displays real-time work status and progress, with support for custom appearances.
Source: IT之家
Read on 88z.aiBaidu AI Cloud has launched Hogee, a one-stop AI platform for short drama creation that automates the entire process from script to final video and includes a digital employee matrix for operations and marketing support.
Source: Hogee
Read on 88z.aiDeepSeek has released a multimodal large model and its technical report on GitHub, introducing a thinking framework based on visual primitives. By converting spatial markers into basic reasoning units, the model significantly enhances spatial reference and deduction capabilities. With a compact and efficient architecture, it performs excellently in counting and spatial reasoning tests, rivaling advanced models like GPT-5.4 and Claude-Sonnet-4.6.
Source: IT之家
Read on 88z.aiAlibaba Tongyi has open-sourced Qwen-Scope, an interpretability toolkit for large models based on sparse autoencoder technology. It covers seven models in the Qwen3 and Qwen3.5 series, enabling targeted output control during inference, data classification, long-tail sample synthesis, and optimization of language mixing and repetitive generation issues.
Source: 千问大模型
Read on 88z.aiThe US Department of Defense announced agreements with seven tech giants, including SpaceX and OpenAI, to accelerate the military's transition to AI-led operations and enhance decision-making advantages across domains. This marks the formal deep integration of top civilian AI technologies into the defense system, promoting their scaled application in military command and intelligence analysis.
Source: 财联社
Read on 88z.aiAlibaba Cloud has officially launched the digital employee platform QoderWake and the Qoder mobile application. Built on a Harness-First architecture, QoderWake features multi-dimensional self-evolution capabilities that consolidate experience into five areas, including memory and skills, addressing the issue of general agents forgetting tasks after completion. The platform currently offers a "Digital Programmer" that enables unattended root cause analysis and code repair within Alibaba, reducin
Source: 阿里云
Read on 88z.aiAnt Baili has officially open-sourced Ling-2.6-1T, a trillion-parameter comprehensive flagship model. Utilizing a hybrid architecture of MLA and Linear Attention, it achieves high intelligence efficiency with minimal token consumption, matching the comprehensive intelligence of GPT-5.4 (Non-Reasoning). The model attains open-source state-of-the-art results on execution benchmarks like AIME26, SWE-bench Verified, and BFCL-V4, demonstrating robust capabilities in agent execution, code generation,
Source: 百灵大模型
Read on 88z.aiTencent Hunyuan has open-sourced the Hy-MT1.5-1.8B-1.25bit model for offline on-device translation, supporting 33 languages and 1,056 translation directions. Compressed to just 440MB, it runs locally on mobile phones without an internet connection. Built on the 1.8B-parameter Hy-MT1.5, it uses Sherry sparse quantization technology (accepted by ACL 2026) and delivers translation quality surpassing mainstream systems like Google Translate.
Source: 腾讯混元
Read on 88z.aiUnitree has launched a new dual-arm humanoid robot with a starting price of ¥26,900. Featuring 15 to 31 degrees of freedom, it supports both fixed and mobile base deployment. The robot is equipped with high-performance hardware, visual and voice interaction systems, and offers 10 TOPS of head computing power. Designed for multi-scenario applications, it supports low-level secondary development to provide a high-performance solution for rapid deployment.
Source: Unitree宇树
Read on 88z.aiSenseTime has officially open-sourced the SenseNova U1 series, a unified model for native understanding and generation. Built on the proprietary NEO-unify architecture, it integrates multimodal understanding, reasoning, and generation within a single framework. The released lightweight version includes 8B dense and A3B MoE specifications, achieving state-of-the-art performance among open-source models of similar scale in various benchmarks.
Source: 商汤科技SenseTime
Read on 88z.aiStepFun releases Step Image Edit 2, achieving performance surpassing 12B-20B parameter open-source models with only 3.5B parameters. It tops the KRIS-Bench lightweight image editing leaderboard and supports generation, editing, and multilingual rendering.
Source: 阶跃星辰
Read on 88z.aiiFlytek has officially launched the Spark X2-Flash large language model and opened its API. The model uses a MoE architecture with 30 billion total parameters, supports a 256K long context window, and was trained on Huawei Ascend 910B domestic computing clusters. It shows significant improvements in agent and coding capabilities, with performance close to trillion-parameter models, token costs less than one-third of mainstream models, and integration with platforms like AstronClaw and Loomy, com
Source: 讯飞开放平台
Read on 88z.aiAnt Group's Bailin has officially open-sourced Ling-2.6-flash, offering multiple precision versions on major platforms. The model has 104B total parameters and previously garnered community attention under an anonymous identity.
Source: 百灵大模型
Read on 88z.aiDeepSeek's multimodal 'Image Recognition Mode' has entered beta testing, with a new entry point visible on the homepage for some users. This feature offers human-like visual understanding, accurately identifying objects, scenes, and visual illusions, while also pinpointing specific locations like Hangzhou's Lingyin Temple through architectural and textual clues.
Source: 智东西
Read on 88z.aiUK AI startup Ineffable has completed an $1.1 billion seed funding round, reaching a $5.1 billion valuation and setting a European record. Founded by former DeepMind chief scientist David Silver, the company is backed by Sequoia and Lightspeed, with participation from Nvidia and Google, aiming to develop autonomous super learning systems that do not require human data.
Source: 智东西
Read on 88z.aiGitHub Copilot announced a shift to pay-as-you-go billing starting June 1, replacing fixed quotas with AI Credits based on token consumption. Base subscription prices remain unchanged, with Pro and Pro+ plans including equivalent credits, while code completion does not consume quotas. The enterprise version supports shared credit pools, and Business and Enterprise customers will receive additional promotional credits from June to August. Annual subscribers will maintain their original billing me
Source: IT之家
Read on 88z.aiXiaohongshu has released its first AI governance principles, establishing three core tenets: supporting AI as a creative aid, requiring labeling of AI-generated content either by users or the platform, and strictly prohibiting AI-driven fraud and infringement. The platform will enhance its detection capabilities and open an appeal channel to protect creator rights and maintain community authenticity and fairness.
Source: 小红书
Read on 88z.aiMicrosoft and OpenAI have revised their agreement, removing exclusive cloud partnership restrictions. While OpenAI products remain prioritized on Azure, services can now be offered to customers via other cloud providers. Intellectual property licensing has shifted from exclusive to non-exclusive, extended until 2032. Microsoft will no longer pay revenue shares to OpenAI, while OpenAI's payments to Microsoft are capped and decoupled from technical progress. Microsoft also gains the right to indep
Source: APPSO
Read on 88z.aiBaidu launched the general agent GenFlow 4.0 at AI DAY, with monthly active users exceeding 100 million. The new version upgrades the Office Agent to support one-click PPT generation, Excel processing, and Word document writing, while enhancing memory center capabilities. It also deeply integrates with OpenClaw for seamless cross-device collaboration and skill migration.
Source: 百度
Read on 88z.aiAlibaba Tongyi Lab has released EAPO, a reinforcement learning framework that optimizes long-text reasoning through an evidence reward mechanism to address model hallucinations. Tests based on Qwen3-30B show its performance surpassing larger mainstream models like GPT-OSS and Claude-Sonnet-4.
Source: 通义实验室
Read on 88z.aiAlibaba QoderWork launches the Expert Suite, featuring ten pre-built kits across six sectors including finance and law, supporting one-click usage and custom encapsulation.
Source: 阿里云
Read on 88z.aiLibTV has premiered the new HappyHorse 1.0 video model, supporting text-to-video, image-to-video, and reference image generation. It produces 15-second 1080P multi-shot narrative videos with features like natural language editing, cinematic quality, smart storyboarding, audio-visual sync, and style restoration, significantly improving character motion, micro-expressions, and dialogue realism.
Source: 哩布哩布AI
Read on 88z.aiAlibaba's Duiyou platform launches the HappyHorse 1.0 video model, supporting scenarios like 3D animation and drone shots, with limited-time discounts as low as 12 Duidou per second.
Source: 堆友
Read on 88z.aiDeepSeek officially announced that the price for input cache hits across all API services has been reduced to one-tenth of the original rate, aiming to lower developer costs and improve cost-effectiveness in long-context scenarios.
Source: IT之家
Read on 88z.aiXiaomi's Robotics Division has open-sourced the full real-device post-training pipeline for Xiaomi-Robotics-0. Built on a pre-trained base, the model achieves high-precision tasks like earphone storage with just 20 hours of task data. The team employed strategies such as asynchronous inference, action prefixes with adaptive weighting, Lambda-type masking, and random masking to address the model's 'laziness effect,' balancing action coherence and visual sensitivity.
Source: 小米技术
Read on 88z.aiLingguang App has launched the 'Experience World Model' feature, bringing world model capabilities to mobile devices for the first time. Users can generate interactive 3D virtual worlds from images and explore them immersively in first-person view.
Source: 灵光App
Read on 88z.aiEmbodied AI company Xingdong Jiyuan has completed a new funding round exceeding $200 million, led by SF Group with participation from Sequoia China, IDG Capital, and CICC. The company has achieved the industry's first product-market fit, deployed solutions in multiple logistics centers with efficiency surpassing humans by 85% in some scenarios, and plans to begin mass delivery of thousands of units in Q2 2026.
Source: 星动纪元
Read on 88z.aiStepFun and Tencent Cloud have formed a strategic partnership to integrate voice large models and multimodal technologies with Tencent Cloud's infrastructure advantages, accelerating the adoption of large models in vehicles. The collaboration focuses on smart cockpits, creating proactive AI interactions featuring wake-word-free operation, continuous dialogue, and emotion recognition, while linking with Tencent Music, video, and map services to achieve a closed-loop in-car service ecosystem.
Source: 阶跃星辰
Read on 88z.aiMango Media and AIsphere have signed a strategic cooperation agreement. Leveraging the Mango Panoramic Workstation, they will integrate AIsphere's AI video large model capabilities to promote the application of technology in audio-video creation, AI film-games, and IP development. The two parties will cooperate in content production, interactive entertainment, and content globalization, exploring industrial applications of AI video in short dramas, animation, variety show visuals, and real-time
Source: 爱诗科技 AIsphere
Read on 88z.aiDeepSeek has released and open-sourced the preview of DeepSeek-V4, featuring Pro and Flash variants that both support a 1M context window. The Pro version matches top closed-source models in performance, while the Flash version offers similar reasoning capabilities at a lower cost. The new architecture significantly reduces computation and memory overhead for long contexts.
Source: DeepSeek
Read on 88z.aiOpenAI has released its new flagship model, GPT-5.5, designed for practical work and agents. Co-designed with NVIDIA GB200/GB300 NVL72 systems, it outperforms previous generations and Claude Opus 4.7 in coding, knowledge work, and scientific research. GPT-5.5 breaks performance bottlenecks, matching GPT-5.4's speed while consuming fewer tokens.
Source: 量子位
Read on 88z.aiiFlytek's AI office assistant Loomy is now fully open for registration without invitation codes. Having completed over 5,000 task loops across more than 20 industries including finance and e-commerce, the latest update introduces natural language control for IM tools, guided PPT creation, and parallel task processing to boost efficiency.
Source: 讯飞开放平台
Read on 88z.aiStepFun has released StepAudio 2.5 ASR, incorporating large language model inference acceleration technology. Built on the ASR+MTP-5 architecture, the model significantly improves speed and reduces costs, achieving SOTA levels in Chinese and English evaluations while supporting transcription of 30-minute long audio clips.
Source: 阶跃星辰
Read on 88z.aiAlibaba Cloud has released JVS Crew, a platform designed to help enterprises embed production-grade AI Agent capabilities into existing applications through an 'integrated' approach. The platform strengthens security and control in enterprise scenarios by providing triple protection for identity, content, and execution, along with managed memory, end-to-end observability, and multi-level cost control.
Source: 阿里云JVS Crew
Read on 88z.aiOpenAI has introduced Workspace Agents in ChatGPT, enabling teams to create collaborative agents for complex tasks. Powered by Codex, these agents handle files, run code, and store memory, operating 24/7 in the cloud with Slack integration. The feature is currently available for Business, Enterprise, Edu, and Teachers plans.
Source: IT之家
Read on 88z.aiAlibaba's Tongyi Qianwen team has open-sourced Qwen3.6-27B, a dense multimodal model with 27 billion parameters. Supporting both multimodal thinking and non-thinking modes, it achieves flagship-level breakthroughs in agent coding capabilities, outperforming the previous open-source flagship Qwen3.5-397B-A17B on benchmarks like SWE-bench and Terminal-Bench. Its dense architecture allows deployment without MoE routing, facilitating broader adoption.
Source: 千问大模型
Read on 88z.aiByteDance has released Seed3D 2.0, achieving state-of-the-art performance in geometry and texture metrics. The model employs a two-stage DiT strategy and a unified PBR architecture, integrating MoE and VLM priors to enhance material realism. It also expands capabilities such as part decomposition, moving 3D generation toward production readiness.
Source: 字节跳动Seed
Read on 88z.aiThe Tencent Hunyuan team has released the Hy3 preview model, marking their first achievement after rebuilding training infrastructure. Featuring a Mixture-of-Experts architecture that integrates fast and slow thinking, the model has 295B total parameters, 21B activated parameters, and supports a 256K context window. As the most intelligent model in the Hunyuan series, it excels in challenging reasoning benchmarks, with significantly enhanced agent and coding capabilities, enabling the one-shot g
Source: 腾讯
Read on 88z.aiOpenAI has open-sourced a privacy filter model that supports local execution to detect and edit personal identifiable information in text.
Source: Filter
Read on 88z.aiStepFun and Qianli Technology have formed a comprehensive strategic partnership to co-develop a native intelligent driving foundation model. StepFun's open-source flagship model, Step 3.5 Flash, has achieved leading results in various benchmarks, backed by its deep layout in Agent foundation models and omni-modal technologies.
Source: 阶跃星辰
Read on 88z.aiAccording to Kechuang Daily, AI startup Anthropic's valuation on pre-IPO equity trading platforms has risen to approximately $1 trillion, surpassing OpenAI's $880 billion. This surge is driven by intense buyer competition for scarce secondary market shares.
Source: 财联社
Read on 88z.aiOpenAI has officially released ChatGPT Images 2.0, its first image generation model with reasoning capabilities. The model excels in accuracy, timeliness, consistency, and visual coherence, generating realistic complex images such as app screenshots, video interfaces, ads, and posters from simple prompts, with significantly improved Chinese text rendering.
Source: 量子位
Read on 88z.aiSpaceX announced a deep partnership with AI coding unicorn Cursor, planning a $60 billion acquisition later this year. If the deal is not executed, a $10 billion cooperation fee will be paid. The two will combine Cursor's products with SpaceX's Colossus supercomputer to build a top-tier coding AI.
Source: 智东西
Read on 88z.aiLoopit's parent company, Yongyue Intelligence, secured a $50 million investment led by Garena, with participation from BlueRun Ventures and others, bringing its year-to-date financing to nearly $100 million. Founded by former Baichuan AI co-founder Chen Weipeng, the app reached the top of the Google Play entertainment chart within two months and received praise from Elon Musk.
Source: 蓝驰创投
Read on 88z.aiAnt Group's Bailling large model team has released Ling-2.6-flash, featuring hybrid attention and sparse MoE architecture. The model achieves a reasoning speed of 340 tokens per second on four H20 cards, with token consumption at only one-tenth of similar products, while reaching state-of-the-art performance on multiple agent benchmarks.
Source: 百灵大模型
Read on 88z.aiSenseAuto has launched Sage, the industry's first complex agent foundation model for vehicles. Built on a MoE architecture, it achieved strong results in PinchBench benchmarks and is deployed on the NVIDIA Orin X platform.
Source: 绝影智能SenseAuto
Read on 88z.aiMoonshot AI releases open-source model Kimi K2.6, strengthening code generation, long-horizon tasks, and agent swarm performance, with benchmark results comparable to leading closed-source models.
Source: 月之暗面 Kimi
Read on 88z.aiTencent has officially launched the beta test for the international version of QClaw, covering regions including the US, Canada, and Singapore. The product emphasizes zero thresholds and no deployment requirements, allowing remote computer control via WhatsApp or Telegram for tasks like tax filing. Notably, 99% of its code was autonomously written by AI in five days.
Source: 腾讯
Read on 88z.aiAlibaba Tongyi Lab announced the official upgrade of AgentScope Tuner, a one-stop automatic optimization engine designed for Agentic AI. It offers capabilities in prompt tuning, model selection, and reinforcement fine-tuning, covering the entire R&D lifecycle.
Source: 通义实验室
Read on 88z.aiAnthropic has launched Claude Design, an AI-native visual design platform powered by Claude Opus 4.7. It features automatic brand system extraction, multimodal input support, and one-click code generation, directly challenging traditional tools like Adobe and Figma, causing a sharp drop in their stock prices.
Source: 量子位
Read on 88z.aiAlibaba Tongyi Qwen releases Qwen3.6-Max-Preview preview, significantly enhancing coding, knowledge, and instruction following, with support for online experience and API calls.
Source: 千问大模型
Read on 88z.aiAlibaba Tongyi has launched Fun-ASR 1.5, an end-to-end speech recognition model supporting 30 languages and seven major Chinese dialects. It features significantly reduced character error rates and specialized ancient poetry recognition. Built on MoE architecture with smart punctuation and text normalization, it is now available on Alibaba Cloud Bailian and ModelScope.
Source: 千问大模型
Read on 88z.aiZhipu AI's AutoClaw has officially launched a self-evolution mechanism and a Skill Store, enabling personalized agent growth by solidifying user preferences. It also introduces the GLM Office Skills suite based on GLM-5.1, supporting intelligent generation and format conversion for office scenarios.
Source: 智谱
Read on 88z.aiAnthropic has officially released Claude Opus 4.7, positioning it as the most powerful large model currently available for widespread use. This upgrade focuses on enhancing complex task execution, high-definition visual understanding, and the stability of long-chain workflows, achieving significant breakthroughs in multiple benchmark tests.
Source: 新智元
Read on 88z.aiAlibaba's Tongyi Qianwen team has open-sourced the sparse MoE model Qwen3.6-35B-A3B. With 35 billion total parameters but only 3 billion activated, it is optimized for agent programming. Its performance significantly surpasses previous generations, rivaling larger dense models. It supports multimodal reasoning and dual-mode switching, excelling in code generation and tool use.
Source: 千问大模型
Read on 88z.aiMiniMax has launched MaxHermes, the world's first cloud sandbox and a self-evolving AI assistant built on Hermes Agent. Its core innovation is a learning loop that automatically extracts reusable skills for continuous iteration, featuring persistent memory and multi-agent parallelism. The product offers zero-threshold deployment, integrates with IM platforms like Feishu, DingTalk, and WeCom, and supports Token Plan consumption deduction.
Source: MiniMax 稀宇科技
Read on 88z.aiTencent has officially released and open-sourced the Hunyuan 3D World Model 2.0. It supports multi-modal inputs to generate complete 3D worlds, exports assets in various formats, and integrates seamlessly with major game engines. The new version uses the WorldMirror 2.0 architecture, significantly improving visual realism and supporting free character exploration and physical collision.
Source: 腾讯混元
Read on 88z.aiAlibaba's ATH Innovation Division has launched HappyOyster, an open-world model product now in beta testing. It supports real-time construction and interaction with directing and wandering modes, and applications are open on the official website.
Source: HappyOyster
Read on 88z.aiStepFun has released the new StepAudio 2.5 TTS voice generation model, featuring global and in-text context control along with zero-shot cloning. Users can precisely adjust voice details via natural language, enabling a shift from mere voice reproduction to creative expression, while supporting high-quality speech generation for any voice without retraining.
Source: 阶跃星辰
Read on 88z.aiHiDream.ai has secured over 500 million yuan in new funding to develop native full-modal world models, enterprise agent products, and expand globally. Its open-source model has exceeded 2 million downloads, serving 30 million users and over 40,000 enterprises across marketing, film, and social media scenarios.
Source: 讯飞创投
Read on 88z.aiNetEase Youdao has launched its first AI knowledge base product, 'Youdao Baoku', positioned as an external super knowledge base that supports document storage, intelligent Q&A, and one-click generation of PPTs, podcast scripts, and research reports.
Source: 网易有道
Read on 88z.aiBaidu Wenxin has officially open-sourced the ERNIE-Image text-to-image model. With only 8 billion parameters, it achieves state-of-the-art open-source performance, matching commercial closed-source models in text rendering and complex instruction following. The model supports multilingual glyph generation, runs on 24GB VRAM, and its resources are available on Hugging Face.
Source: 百度文心
Read on 88z.aiAlibaba's ATH Business Group has launched Meoo, its first AI development tool. Integrating four leading models including Qwen, Kimi, GLM, and MiniMax, along with core Alibaba Cloud services, it allows users to generate full-stack code via natural language without coding skills and deploy with one click.
Source: 阿里云
Read on 88z.aiWorld Labs, led by Fei-Fei Li, has open-sourced Spark 2.0, a 3D Gaussian Splatting rendering engine that supports streaming over 100 million splats via WebGL2 to various devices.
Source: 量子位
Read on 88z.aiMeta has entered a five-year partnership with Broadcom, initially committing over 1GW of compute power for custom AI chips, aiming to provide personal superintelligence to billions. Meta plans to launch four generations of self-developed MTIA chips within two years and has established a Superintelligence Lab led by the founder of Scale AI as Chief AI Officer.
Source: 新智元
Read on 88z.aiModelBest has launched Lantay, a professional document agent workspace, now in public beta. Inspired by Cursor's design philosophy, it features a unified layout and Vibedocing interaction paradigm. Supporting 16 file formats and bulk processing, it targets heavy document users in rigorous industries like law and finance.
Source: 面壁智能
Read on 88z.aiVolcengine has officially launched the Seedance 2.0 API, supporting multi-modal input and editing for text, images, audio, and video. It establishes portrait copyright safety standards and is widely used in film, advertising, and autonomous driving.
Source: 火山引擎
Read on 88z.aiMiniMax Agent desktop releases two updates: Pocket enables direct computer task execution via IM apps like Feishu and WeChat; Computer Use allows the agent to control mouse and keyboard like a human, managing local software and system settings.
Source: MiniMax 稀宇科技
Read on 88z.aiMiniMax has officially open-sourced MiniMax M2.7, a new self-evolving AI model deeply iterated by AI itself. It excels in building complex Agent systems and performs well in software engineering and office scenarios. On its first day of release, it achieved compatibility with major domestic and international chip and inference platforms.
Source: MiniMax 稀宇科技
Read on 88z.aiMiniMax releases Music 2.6 model with new Cover adaptation feature, optimized performance for traditional Chinese instruments, mid-low frequencies, and vocals. First packet latency is under 20 seconds, supporting precise instruction control, with three open-source Music Skills.
Source: MiniMax 稀宇科技
Read on 88z.aiAnuttacon, the AI company founded by MiHoYo co-founder Cai Haoyu, has officially launched its first video generation model, LPM 1.0. With 17 billion parameters, the model supports real-time generation of AI character videos that can speak, sing, listen, and express emotions. Its key advantage lies in maintaining identity consistency for videos lasting 22 to 48 minutes. Technically, LPM 1.0 utilizes a diffusion Transformer architecture combined with a causal flow generator, effectively addressing
Source: 量子位
Read on 88z.aiAlibaba Tongyi Lab has released the VimRAG framework, replacing traditional linear context concatenation with dynamic directed acyclic graphs. By integrating branch trial-and-error, visual energy allocation, and graph-guided optimization, it enhances retrieval and reasoning across text, images, and video. Tests on the Qwen3-VL-8B model show an accuracy of 50.1%, significantly outperforming conventional approaches.
Source: 通义实验室
Read on 88z.aiJimeng AI has launched Octo, an intelligent creative assistant centered on the Vibe Create philosophy, enabling unrestricted free creation. The tool is now available for beta testing on the official website, aiming to help users easily realize ideas through pure conversational experiences.
Source: Octo
Read on 88z.aiMeta's Superintelligence Lab has released Muse Spark, its first native multimodal model, which achieved a significant score jump in Artificial Analysis benchmarks and drove strong market performance. The model features visual chain-of-thought and multi-agent orchestration capabilities, demonstrating excellence in specific vertical domains and boosting the company's stock price.
Source: 新智元
Read on 88z.aiAI companion robot company Mengyou Intelligent ropet has announced the completion of its Series A financing, raising over $10 million. Investors include the Beijing Artificial Intelligence Industry Investment Fund and FreeS Fund. The company focuses on desktop AI pets priced at approximately 2,000 yuan, with nearly 20,000 units shipped to date. Data shows a 90-day retention rate of 80% to 90%, with core users interacting for over two hours daily. The product emphasizes building emotional connect
Source: 极客公园
Read on 88z.aiByteDance's Seed team has released Seeduplex, a full-duplex voice large model now fully deployed on the Doubao App. Using a 'listen-while-speaking' architecture, the model achieves simultaneous listening and speaking, significantly reducing error replies and interruptions while cutting stop-judgment latency. It improves conversation fluency and user satisfaction, marking the first large-scale implementation of full-duplex voice technology.
Source: 字节跳动Seed
Read on 88z.aiClaude has launched Claude Managed Agents, an enterprise-grade managed service providing production-ready infrastructure to help developers quickly build and deploy agents. Adopted by companies like Notion, it reportedly boosts development speed tenfold and is now in public beta with usage-based pricing.
Source: 量子位
Read on 88z.aiMiniMax has released MMX-CLI, a command-line tool built for AI Agents that enables direct access to its multimodal models within environments like Claude Code, allowing automated workflows from research to video production without additional adaptation.
Source: MiniMax 稀宇科技
Read on 88z.aiTencent has launched QBotClaw, China's first browser-based AI agent, integrated directly into QQ Browser for immediate, download-free access. The assistant handles complex tasks such as research, copywriting, trend monitoring, price comparison, and desktop organization via simple natural language commands. It also enables remote computer control through WeChat Clawbot. Featuring browser context memory and high-precision web recognition, QBotClaw is currently available on Mac, with a Windows vers
Source: QQ浏览器
Read on 88z.aiZhipu has released its open-source flagship model, GLM-5.1, claiming it to be the world's most powerful open-source model. Capable of working independently for over eight hours, it can autonomously handle complex engineering tasks. The model ranks third globally and first among domestic models in coding benchmarks like SWE-Bench Pro, surpassing GPT-5.4 and Claude Opus 4.6. Tests demonstrate its ability to build complete Linux desktop systems and significantly optimize vector database and ML work
Source: 智谱
Read on 88z.aiAIsphere has launched PixVerse C1, the world's first large model designed for the film and television industry. It supports various generation modes and intelligent storyboarding, producing videos up to 15 seconds in 1080P resolution. The model achieves breakthroughs in multi-character scheduling, character consistency, and industrial-grade visual effects, enabling one-click video production.
Source: 爱诗科技 AIsphere
Read on 88z.aiAnthropic has released the preview version of Claude Mythos, which outperforms Opus 4.6. Due to high security risks, access is restricted to giants like Amazon, Apple, and Google, as well as over 40 infrastructure companies via Project Glasswing for defensive security hardening.
Source: 量子位
Read on 88z.aiFormer OpenAI scientist Andrej Karpathy introduced LLM Wiki, a scheme proposing that AI agents automatically build personal knowledge bases from shared idea files in the Agent era. The system features a three-layer architecture of data, Wiki, and Schema, forming a closed loop to support continuous knowledge accumulation for research and enterprise knowledge management.
Source: 机器之心
Read on 88z.aiModelBest has completed a new round of financing worth hundreds of millions of RMB, co-led by Shenzhen Capital Group and Huichuan Industrial Investment. Recognized as a 2026 Chinese Unicorn, the company focuses on on-device large models. Its MiniCPM series has exceeded 24 million downloads and achieved scaled deployment in automotive, smartphone, and smart home sectors.
Source: 面壁智能
Read on 88z.aiOpenBMB has released VoxCPM2, an open-source speech synthesis model with 2 billion parameters. It supports 30 languages and Chinese dialects, delivering 48kHz studio-quality audio. The model introduces Voice Design for creating new voices via text descriptions, alongside controllable and ultimate cloning modes. Trained on 2.36 million hours of data with a real-time factor of 0.13, it is licensed under Apache-2.0 for commercial use.
Source: VoxCPM2
Read on 88z.aiByteDance's Coze 2.5 has officially launched, introducing the 'Agent World' ecosystem. The new version equips AI Agents with independent cloud computers and cloud phones, supporting 24/7 autonomous background task execution; it integrates video creation, programming CLI, and expert skills in industries like law and finance; and introduces a long-term memory system to enable cross-platform personalized collaboration.
Source: 扣子Coze
Read on 88z.aiMicrosoft has released MAI-Transcribe-1, claiming it is the world's most accurate AI transcription model with a 3.9% word error rate. Supporting 25 languages, it outperforms Whisper-large-v3 and Gemini 3.1 Flash. Available on Microsoft Foundry, it offers batch transcription at 2.5 times the speed of Azure Fast services for $0.36 per hour, though real-time transcription and speaker separation are not yet supported.
Source: IT之家
Read on 88z.aiGoogle has released the Gemma 4 series of open models, featuring E2B, E4B, 26B MoE, and 31B Dense versions. Designed for advanced reasoning and agent workflows, this series is touted as the most powerful in its class, supporting multimodality and long contexts under an Apache 2.0 open-source license.
Source: 谷歌开发者
Read on 88z.aiXiaomi has launched the Xiaomi MiMo Token Plan subscription, adopting a unified Credit billing system and removing the 5-hour token limit, with four tiers available to support mainstream AI development tools.
Source: Xiaomi MiMo
Read on 88z.aiAI video platform OiiOii has officially opened registration, integrating the full-power Seedance 2.0 model for high-quality video generation without queues. The update introduces multi-panel storyboard previews, a free canvas mode, and scene asset features, along with new styles like realistic human portraits, significantly boosting creation efficiency and visual consistency.
Source: OiiOii AI
Read on 88z.aiAlibaba Tongyi Lab releases the Wan2.7-Video model, supporting full-modal input and instruction-based video editing, with capabilities for subject consistency control, storyboard segmentation, and intelligent plot design.
Source: 通义实验室
Read on 88z.aiZhipu has released GLM-5V-Turbo, a multimodal coding foundation model that natively integrates visual and text capabilities. It supports generating runnable code directly from design drafts or screenshots, with a context window of 200k. The model leads in multimodal coding and GUI Agent benchmarks and is deeply adapted to frameworks like Claude Code and AutoClaw, enabling AI agents with visual perception.
Source: 智谱
Read on 88z.aiTongyi Lab releases Qwen3.6-Plus, accessible via Alibaba Cloud Bailian API. The model focuses on coding agent capabilities, supports million-level context, enhances multimodal reasoning, and leads in code generation, repair, and front-end development tasks.
Source: 通义实验室
Read on 88z.aiOpenClaw has officially launched the ClawHub mirror site, with infrastructure support from ByteDance's BytePlus and Volcano Engine, aiming to resolve slow access to the skills market in China. The site offers a complete Chinese interface and data synchronization, continuously curating high-quality community Skills. Users can access it directly via the specified domain and seamlessly invoke various Agent skills within OpenClaw.
Read on 88z.aiBilibili's self-developed AI tool updream has entered beta testing, offering invited creators lightweight intelligent services including inspiration generation, smart editing, and project management.
Source: updream
Read on 88z.aiOpenAI closed a record $122 billion financing round, reaching an $852 billion valuation with lead investors including Amazon and Nvidia. Due to high costs and low retention, Sora was shut down as the company pivots to text, code, and enterprise services to prepare for an IPO.
Source: APPSO
Read on 88z.aiTongyi Lab has released the Wan2.7-Image model, focusing on human realism, text stability, and color accuracy. The model supports deep customization of virtual avatars to avoid homogeneous faces, stably generates images with ultra-long text including multiple languages, tables, and formulas, and introduces a Hex value palette for precise control of brand colors.
Source: 通义实验室
Read on 88z.aiThe Meituan LongCat team has released the LongCat-AudioDiT model, achieving state-of-the-art performance in zero-shot voice cloning. The model performs diffusion generation directly in the waveform latent space, discarding traditional Mel-spectrogram intermediate representations to prevent information loss. It also introduces Dual Constraint Alignment and Adaptive Projection Guidance techniques to address training-inference mismatches and mitigate oversaturation.
Source: 龙猫LongCat
Read on 88z.aiAIsphere announced three major updates for PixVerse on the second day of its Lightning Release Week: a Quick App, a Team Edition, and a CLI tool. The Quick App debuts with 'Ad Master,' enabling rapid conversion of product images into ad videos. The Team Edition supports collaboration for 2 to 100 users, offering shared credits, four-level permission management, and independent spaces. The CLI tool allows developers to invoke video generation capabilities via command line, facilitating integratio
Source: 爱诗科技 AIsphere
Read on 88z.aiKunlun Wanwei has launched the Tiangong Short Drama Workbench, integrating four major video models and multi-agent technology to provide a one-stop industrial creation process from script to final film.
Source: 昆仑万维集团
Read on 88z.ai