Mistral releases open weights for improved Mistral Large 4 (Le Chonk), a 1T-parameter multimodal model, after a public preview earlier in October

Mistral has published the weights of an updated Mistral Large 4, a 1-trillion-parameter model that handles images as well as text. It was in public preview earlier in October and has had about three weeks of extra reinforcement learning since. It now has three reasoning modes (Low, Medium, High by default), and Low replaces the old "None" mode. Mistral reports 19.7% on Cognition's private FrontierCode 2.0, 48% on the Artificial Analysis Coding Agent Index (roughly level with GLM-5.3, using 47% fewer tokens), and 41% on Terminal-Bench 4.0 (up 13 points from the preview). All of these figures come from Mistral's own blog.

Developers and enterprises can now download and self-host a very large open-weight model that Mistral says is competitive for coding. That matters most to organisations that need data control or want a non-US, non-Chinese option. A 1-trillion-parameter model needs serious hardware, so in practice this suits well-resourced teams and hosting providers, not small shops. The lower token use, if it holds up, would cut running costs for coding agents. The gains over the preview are incremental, so this is a refinement of a known model, not a new capability step.

"Best open-weight model for coding" rests on Mistral's self-reported results, one of which is a private benchmark (FrontierCode 2.0) that outsiders cannot inspect, and the comparisons are against models Mistral chose. The "trusted European foundation for sovereign AI" line is positioning, not a measured property, and the 96.6% refusal rate on malicious cyber requests says nothing about how it performs against determined attackers. The human-preference win over three rivals is also Mistral's summary, and it still lost to Claude Opus 5.5.

1. 01Independent runs of the Artificial Analysis Coding Agent Index and Terminal-Bench 4.0 on the released weights, to see whether the 48% and 41% figures hold and whether the token savings are real.
2. 02Real-world uptake over the next 4 to 6 weeks: availability on major inference hosts, quantised versions people can actually run, and published cost per task against GLM-5.3 and Kimi K3.
3. 03Third-party red-teaming of the cyber behaviour, in particular whether the 96.6% refusal rate survives jailbreaks and whether defensive tasks are handled without over-refusal.