TL;DR

Mistral AI released the weights of Mistral Large 4, nicknamed “Le Chonk”, on 29 October, weeks after the 1-trillion-parameter multimodal model entered public preview. The company says extra reinforcement learning has made it the strongest open-weight model for coding. It also says monitoring of live traffic showed the model could give real attackers an uplift, and that it is tightening guardrails in response.

The claims

All performance figures come from Mistral and have not been independently checked. It says three further weeks of reinforcement learning after the preview put Large 4 ahead of other open-weight models on FrontierCode and Terminal-Bench. On Terminal-Bench 4.0 its score rose 13 points to 41% while output tokens fell by 30%. Artificial Analysis’s coding-agent index gives it 48%, which Mistral calls effectively level with China’s GLM-5.3 while it uses 47% fewer tokens on average.

Developers can now choose low, medium or high reasoning effort to trade accuracy against cost. API pricing is $0.70 per million input tokens and $2.10 per million output tokens, with cached input at $0.07.

Open weights, with a catch

The model ships under what Mistral calls its Ecosystem Open-Weight licence, a modified MIT licence. Anyone can use, adapt and sell products built on it, but companies offering it as a hosted model service need a separate commercial licence first. Mistral pitches the release as sovereign AI, stressing that Large 4 was built from the ground up in Europe and can run entirely on a customer’s own servers. Microsoft is the first cloud to carry it, through Foundry.

The cyber question

On Artificial Analysis’s cyber index, which tests finding, reproducing and patching vulnerabilities, Mistral reports 51%. Its own frankness on risk is notable: the model declined black-box attacks on live targets in its traffic, yet Mistral says it saw potential for uplift in genuine attacks. A specialised security variant goes only to vetted partners, Orange Cyberdefense among them.

The timing is pointed. Reuters reported the same day that GLM-5.3, the model Mistral benchmarks itself against, completed 3 of 13 simulated network attacks in tests by security firm Irregular, sharpening concern about capable models anyone can download.

Looking Forward

In our view, for UK organisations weighing where to run AI, an open-weight frontier model trained outside the US and China widens the self-hosting options, and the licence lets most businesses deploy it without paying Mistral. We would weigh the cyber caveat as heavily as the coding scores: once weights are public, guardrail changes Mistral makes to its own service cannot be applied to copies others run. Mistral says a fuller technical report will follow.