Mistral releases weights for trillion-parameter Mistral Large 4, targeting cybersecurity

What happened

Mistral has released weights for Mistral Large 4 (ML4), its trillion-parameter model introduced earlier this month, enabling businesses and governments to download and deploy it themselves for cybersecurity work. Mistral claims performance comparable to GLM-5.3 while using 47 percent fewer tokens per task, plus the highest ranking among open models on Artificial Analysis’ Coding Agent Index. The weights are available on Hugging Face, and the model is also accessible through Mistral’s API.

Read next

• More changes in today’s briefingGoogle DeepMind announces Gemini 4 Carbon with cybersecurity and long-horizon capabilities

Reported by 6 sources

The story so far

Mistral AI released improved weights for Mistral Large 4, also called Le Chonk, making the multimodal model available for download and self-deployment, with launch-day inference support from NVIDIA and vLLM. Mistral says use, modification, fine-tuning and commercialization are free, but businesses operating Model-as-a-Service offerings need a separate license before commercial use. The Decoder describes the exception more narrowly as applying to companies offering Large 4 as a commercial cloud service, and reports that access requires review and contractual restrictions. The model is also available through Mistral’s API. The release targets agentic coding and cybersecurity. Artificial Analysis lists 1.05 trillion total parameters and 52 billion active parameters, and ranks it narrowly first among open-weight models developed outside China. Mistral reports a 48% Artificial Analysis Coding Agent Index score while using 47% fewer tokens per task than GLM-5.3, and first place among open models on Cognition’s FrontierCode. Those coding results have a different scope from the overall Artificial Analysis score of 26 reported by The Decoder, which places it behind several Chinese rivals. NVIDIA provided NVFP4 weights on Hugging Face and an optimized Dynamo configuration for distributed inference. vLLM added support for FP8 and NVFP4 checkpoints with a built-in DSpark drafter, and describes Large 4 as supporting a 1-million-token context.