Paris-based Mistral AI has released the full open weights of Mistral Large 4, the trillion-parameter flagship it nicknames “Le Chonk” — meaning anyone can now download the model and run it on their own servers.
The release, announced on October 29, follows the model’s public preview on October 6. Mistral says three more weeks of reinforcement learning since the preview lifted its score on Terminal-Bench 4.0, a coding benchmark, from 28.3% to 41%.
Large 4 is a mixture-of-experts model: roughly 1 trillion parameters in total, with about 52 billion active for each token it generates, a context window of one million tokens, and text and image input. In plain terms, it is a very large model that only wakes up a small slice of itself for each word — which is why it can be both huge and comparatively cheap to run.
Mistral pitches it first at coding and cybersecurity. The company reports 19.7% on Cognition’s FrontierCode 2.0 benchmark, the top open-weight score there, and 48% on the Artificial Analysis Coding Agent Index — effectively tied with China’s GLM-5.3, while using 47% fewer tokens per task. On cyber defence, it reports 51% on the Artificial Analysis Cyber Index and says the model refused 96.6% of malicious cyber requests in its tests.
There is, however, a sharper edge. Mistral says it saw “a potential for uplift to conduct real attacks,” and a more aggressive penetration-testing version stays restricted to vetted partners. The public model is the defensive one; the offensive one is not being handed out.
The licence is where the word “open” gets qualified. The weights are published under Mistral’s Ecosystem Open-Weight Licence, a modified MIT licence: free to use, modify, fine-tune and commercialise — with one carve-out. Businesses that offer the model itself as a hosted service must buy a separate commercial licence from Mistral first. Self-hosters, fine-tuners and app builders face no such obligation. Through Mistral’s own API, access costs $0.70 per million input tokens and $2.10 per million output tokens.
Why this matters
Open-weight releases are the counterweight to the closed AI race. A closed model can only be rented; an open-weight model can be inspected, adapted and run inside a hospital, a bank or a government data centre that would never send its data to someone else’s cloud. That is why Mistral frames this as a sovereignty play, trained on European infrastructure. The catch is the same as ever: the more capable the open model, the more the debate about what it can do in the wrong hands moves from theory to practice. Large 4 is now the test case.
That contrasts with closed offerings such as Gemini, which are typically rented rather than downloaded and run locally.
