Mistral Releases Weights for 1-Trillion-Parameter Mistral Large 4

October 30, 2026

French AI lab Mistral has released the open weights for Mistral Large 4, codenamed "Le Chonk", a 1-trillion-parameter natively multimodal model that the company says leads open-weight benchmarks in coding and defensive cybersecurity [1].

The model entered public preview earlier in October 2026. Three weeks of additional reinforcement learning pushed it to the top of open-weight coding benchmarks including FrontierCode and Terminal-Bench, while matching the best open-weight models on general reasoning tasks [1].

Mistral Large 4 offers three reasoning modes, letting developers balance cost and performance depending on the workload. The company reports substantial gains in reasoning efficiency, meaning better performance at lower inference cost, a combination that matters for anyone self-hosting [1].

Trained entirely on European infrastructure with clear lineage from inception to release, Mistral positions the model as a foundation for "Sovereign AI", the idea that organizations should own, customize, and control their AI rather than depending on foreign cloud providers [1].

The release puts pressure on other open-weight labs. A 1-trillion-parameter model competitive with frontier closed models on coding, available for self-hosting, is a meaningful shift. The cybersecurity focus is also notable. Mistral claims the model excels at defensive security tasks, which is a narrower claim than general capability but one that matters for enterprise adoption [1].

The name "Le Chonk" is, admittedly, very French internet humor. The model is anything but chonky in the ways that count. It is fast, efficient, and built for developers who want to run it on their own hardware [1].

Sources

[1] Mistral AI, "Mistral Large 4 (Le Chonk) is now open weight"