88Z AI / NEWS
The news, the context, and what changes.
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.
As 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.
NVIDIA 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.
AllenAI 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.
Google 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.
iFlytek 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.
Fei-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.
Perplexity 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.
For 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.
xAI 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.
Anthropic 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.
Looki 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.
