MiniMax released the M2 model, focusing on new methods for agent alignment and generalization. It proposes an 'interleaved thinking' mechanism to handle long-horizon tasks and perturbations in external tool outputs. The release notes that true generalization lies in adapting to perturbations across the entire operational space (including tool sets, system prompts, user prompts, environments, and tool responses), rather than merely expanding the number of tools. It also emphasizes the critical role of complete conversation history (including thinking steps) for performance.
OpenAI has released the OWL architecture as the underlying technology for the ChatGPT Atlas browser, achieving Chromium decoupling, fast startup, a rich UI, and agentic browsing based on ChatGPT.
Google Research published a paper at UIST’25 proposing the StreetReaderAI prototype system, which integrates an AI Describer and an AI Chat subsystem supported by Gemini.
A Hugging Face blog analysis points out that U.S. export controls on high-end AI chips since 2022 have not curbed China's AI development but instead accelerated the deployment of domestic chips such as Huawei Ascend and Cambricon.
NVIDIA Isaac for Healthcare v0.4 releases an SO-ARM surgical assistant workflow for medical robots, supporting synthetic data generation in simulation, hybrid training of GR00T N1.5 policies, native fine-tuning via LeRobot 0.4.0, and deployment to SO-ARM101 hardware; 93% of training data is simulation-generated, offering dual-mode teleoperation (keyboard/hardware), TensorRT conversion, and an end-to-end Sim2Real pipeline.
OpenAI has released gpt-oss-safeguard, an open-weight reasoning model for safety classification that supports developers in applying and iterating on custom policies.
NVIDIA Isaac for Healthcare v0.4 released an end-to-end workflow for a surgical assistant robot based on SO-ARM, supporting hybrid data collection from simulation and real environments, GR00T N1.5 post-training, and real-world deployment; the article includes full command-line examples, hardware requirements (SO-ARM101 dual-camera robotic arm, RT Core GPU), keyboard teleoperation control keys, and LeRobot v0.4.0's native training support for GR00T N1.5.
IBM released the Granite 4.0 Nano series models, including 4 instruction-tuned models and their corresponding base models, with parameter counts of ~350M and ~1.5B, adopting either hybrid-SSM or traditional Transformer architectures.
In a blog post, Hugging Face proposed the 'voice consent gate' mechanism, which requires speakers to read aloud dynamically generated consent statements containing the model name (e.g., 'I give my consent to use the <MODEL> voice cloning model with my voice') in real time.
The official Hugging Face blog announced that the streaming=True feature in the datasets library has undergone major optimization, reducing startup requests by 100 times, speeding up data parsing by 10 times, and achieving 2x sample throughput. It supports Prefetching for Parquet and configurable buffering; these improvements have been implemented in nanoVLM training, and a streaming-ready FineVisionMax dataset has been pre-built.
Hugging Face officially released huggingface_hub v1.0, marking the entry of this open-source Python library into a mature stage; the new version replaces requests with httpx.
SGLang v0.5.4 has been officially released, featuring several key updates: the launch of Model Gateway v0.2; new support for Nemotron, DeepSeek OCR, and Qwen3-Omni.
The author proposes that AI can serve as a programming collaborator that does not write code. Specifically, this includes: 1) acting as an investigative coding partner to obtain analytical clues for specific issues (such as abnormal function results); 2) acting as a resident bug hunter to troubleshoot error causes under clear instructions, requiring precise input/output and execution paths; 3) acting as a cross-language optimization expert to suggest SQL optimizations such as index adjustments by combining EXPLAIN ANALYZE output with database structure; 4) acting as a personal security reviewer to check for potential vulnerabilities in code changes based on preset security standards; 5) acting as a code reviewer and documentation enhancement tool to identify test gaps, accessibility issues, circular dependencies, etc., and generate acquisition-ready documentation suggestions.
Mistral AI 发布 AI Studio，一个专为将AI从原型推进到生产环境设计的平台，包含 Observability、Agent Runtime 和 AI Registry 三大支柱；它支持内置评估、可追溯反馈闭环、全生命周期资产版本管理、混合/私有部署及企业级治理能力。该平台基于 Mistral 自身大规模AI系统运营经验构建，现开放私测。
Hugging Face officially released LeRobot v0.4.0, introducing scalable Datasets v3.0 (supporting OXE-level data >400GB, chunked episodes, and Parquet metadata).
OpenAI has acquired Software Applications Incorporated, the developer of Sky, a natural language interface for the Mac platform. The two parties will integrate Sky's deep macOS capabilities into ChatGPT to enhance AI intuitiveness, context awareness, and operability.
Hugging Face and Meta announced the joint launch of OpenEnv Hub—a shared open-source center for agent environments—and simultaneously released the OpenEnv 0.1 specification (RFC).
OpenAI has launched the Company Knowledge feature, which brings context from enterprise applications into ChatGPT to provide precise, business-specific answers. It supports explicit citations, security and privacy safeguards, and administrative controls. The feature is currently available to Business, Enterprise, and Edu users.
The Sentence Transformers project has been officially transferred from the UKP Lab at Technical University of Darmstadt in Germany to Hugging Face, with Tom Aarsen continuing to lead maintenance. The project remains under the Apache 2.0 open-source license and supports training of Dense/Cross/Sparse/multimodal embedding and reranking models. Over 16,000 models are hosted on the Hugging Face Hub, serving over one million monthly active users.
Hugging Face has partnered with VirusTotal to continuously scan files in all 2.2 million+ public model and dataset repositories on the Hugging Face Hub starting today; when users access repository or file pages, the Hub automatically compares file hashes and returns VirusTotal's detection status, number of detections, and associated threat intelligence, without uploading raw file content throughout the process.
Hugging Face AI Sheets is an open-source no-code tool that now includes visual support: it can extract image information (such as receipts and documents) directly within spreadsheets, generate images from text, and edit existing images.
Hugging Face published a blog post titled "Supercharge your OCR Pipelines with Open Models," which systematically introduces modern open-source OCR models (such as OlmOCR-2.
Hugging Face and Intel collaborated to benchmark OpenAI's open-source MoE model GPT OSS-120b on Google Cloud C4 (based on Intel Xeon 6 Granite Rapids).
PyTorch 2.9.0 has been officially released. The minimum Python version is raised to 3.10, libtorch ABI introduces unstable updates, and torch.compile enables graph break error control by default.