Briefs · Taipei time (GMT+8)

Mistral releases Le Chonk (Mistral Large 4) weights for coding agents and cybersecurity

一句重點:Teams considering self-hosted open-weight models for coding agents now have another option. NVIDIA describes Le Chonk as a 1T-parameter model.

發生什麼事

@MistralAI announced a new, improved version of the Le Chonk weights for agentic software development and cybersecurity. Mistral says Le Chonk ranked first among open models in Cognition’s FrontierCode evaluation (🗣 self-reported). The Hugging Face Blog also published a post that day; its title identifies Le Chonk as Mistral Large 4 and places the weights on Hugging Face.

@NVIDIAAI describes Le Chonk as a multimodal MoE model with 1T parameters, trained from scratch using 3,800 Grace Blackwell GPUs. NVIDIA says an NVFP4 version was available on Hugging Face at launch, alongside Dynamo support for distributed inference.

01Sources

  • @MistralAI: Mistral’s official account; posted 2026-10-29 at 14:21 UTC (22:21 Taipei); announced the updated Le Chonk weights, intended uses and its FrontierCode claim
  • Welcome Le Chonk: Mistral Large 4 on Hugging Face: Hugging Face’s official blog; the title identifies Le Chonk as Mistral Large 4 and places it on Hugging Face
  • @NVIDIAAI: NVIDIA AI’s official account; model scale and architecture, training, launch-day NVFP4 and Dynamo support