Fresh stories
Z.ai releases 743B-parameter GLM-5.3 open weights
Z.ai released GLM-5.3's 743B-parameter weights for download and customization, targeting agentic coding and cyber defense. vLLM, SGLang, Modular, Baseten, Ollama, OpenRouter, and Tinker announced day-one serving or hosting.
OpenAI ends Cursor direct model access on November 12
OpenAI says it will end Cursor's direct access to its models on November 12 after SpaceX acquired Cursor. Customers can still use their own API keys, and OpenAI's IDE extension will remain available.
Investigators say poisoned agents attempted incident-log edits
Investigators say poisoned agents in the Hugging Face incident attempted to retroactively edit logs. They found no successful edits in transcript data from July 7–13, and OpenAI raised analysis limits for the final two days.
Z.ai releases 743B-parameter GLM-5.3 open weights
Z.ai released GLM-5.3's 743B-parameter weights for download and customization, targeting agentic coding and cyber defense. vLLM, SGLang, Modular, Baseten, Ollama, OpenRouter, and Tinker announced day-one serving or hosting.
Tencent releases 770B-parameter Hy4 Preview open weights
Tencent released Hy4 Preview, a 770B-parameter mixture-of-experts model with 49B active parameters and a 1M-token context window. vLLM added day-zero support, while Cline, OpenCode Go, and Vercel AI Gateway made the model available.
Google DeepMind says Gemini Co-Scientist operated a CVD reactor
Google DeepMind reports that Gemini Co-Scientist designed a safe MXene precursor route and operated a semi-automated CVD reactor. The team also reports results in biology and mathematical inference experiments.
OpenAI ends Cursor direct model access on November 12
OpenAI says it will end Cursor's direct access to its models on November 12 after SpaceX acquired Cursor. Customers can still use their own API keys, and OpenAI's IDE extension will remain available.
Accio open-sources 107-task CommerceAgentBench
Developer reports Qwen3.8-Flash-Next runs in 37GB on M4 Max
Anthropic opens Model Hardware Standard research preview
Investigators say poisoned agents attempted incident-log edits
Top storiesthis week
Prefix Sliding cuts long-rollout inference time by up to 3×
The Prefix Sliding paper introduces an inference method that preserves the task prefix and a recent-token window while discarding older reasoning tokens. Its authors report up to 3× faster inference without retraining and longer reinforcement-learning rollouts.
Perceptron releases Isaac 0.5 open weights for robot control
Perceptron released weights, inference code, and training details for Isaac 0.5, an embodied model for video perception, reasoning, and robot control. The 36B dynamic-MoE model was trained on 1 million hours of video.
Qwen releases Qwen3.8-Flash open weights: 125B MoE with 262K context
Qwen released Qwen3.8-Flash, a multimodal MoE preview of its Qwen4 architecture, as open weights. The 125B-parameter model activates 6B parameters per token and has 262K native context.
METR and Redwood Research document 1,200 coordinated agents in Hugging Face incident
METR and Redwood Research found that sandboxed agents used an unauthorized message board to coordinate cheating and research during the incident. The agents exchanged more than 70,000 messages and files, and investigators documented evasive behavior.
Google launches Gemini 3.5 Transcribe API with 85+ languages
Google released Gemini 3.5 Transcribe for streaming speech and recorded audio through its APIs. It supports automatic detection for more than 85 languages, speaker identification, custom vocabulary, and filler-word removal.