Mistral releases open weights for Mistral Large 4 (Le Chonk)

Mistral AI has published the full weights of Mistral Large 4, codenamed Le Chonk, a 1 trillion parameter multimodal model that entered public preview earlier in October. After three more weeks of reinforcement learning, Mistral says the model now leads open-weight models on coding and cybersecurity benchmarks while using far fewer tokens per task, lowering inference costs. It supports three reasoning modes (low, medium, high) so developers can balance speed, cost, and accuracy.

Mistral's own benchmark numbers include a 19.7% score on Cognition's FrontierCode 2.0 and 41% on Terminal-Bench 4.0, both said to be the best among open-weight models. The company also reports a 96.6% refusal rate on malicious cybersecurity requests, which it calls the highest among open models, while still handling legitimate defensive security work. Several cybersecurity vendors, including CrowdStrike, Cloudflare, and Rubrik, provided supportive quotes about testing the model.

Artificial Analysis, an independent benchmarking group, measured the model differently: 15% on FrontierCode 2.0 and an overall Intelligence Index score of 26, alongside a price of $0.70/$2.10 per million input/output tokens and a 1M token context window. The weights are released under Mistral's own MEOW license, which requires a separate commercial license to run a model-as-a-service business.

Why it matters

It gives developers another capable, self-hostable open-weight coding model, though the performance claims vary depending on who measured them.

Source: Mistral AI