Article 343: Mistral Put a Trillion-Parameter Model on the Open Internet for Half Price. The Interesting Part Is What It Refused

Mistral released the weights of Mistral Large 4 - codenamed Le Chonk - on Thursday. One trillion parameters, mixture of experts with 52B active, natively multimodal, trained from scratch on European infrastructure. Downloadable by anyone. The API now charges $0.70 per million input tokens and $2.10 per million output.

Three weeks ago this model was in public preview. Three weeks of additional reinforcement learning moved it to the top of the open-weight coding table.

The numbers worth keeping

  • Terminal-Bench 4.0: 41%, up 13 points from preview, leading open-weight models, with 30% fewer output tokens.
  • Artificial Analysis Coding Agent Index: 48%, effectively tied with GLM-5.3, using 47% fewer tokens per task.
  • Cognition's FrontierCode 2.0: 19.7%. That benchmark asks whether a repository maintainer would actually merge the agent's change. Nineteen point seven per cent is the honest number and the useful one.
  • Surge Code blind pairwise: preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash; behind only Claude Opus 5.5.

The token-efficiency line is the one that matters commercially. A model that matches a rival's score while spending half the tokens is a different product at scale, not a marginally better one.

The refusal number that deserves attention

Mistral says the model refused 96.6 per cent of malicious hacking requests in its own tests. Then it says the quiet part: watching real traffic suggested it could still help with real attacks, so it is tightening the guardrails.

That is an unusually candid disclosure in an open-weights launch, and it is the correct one. The weights are already out. Anyone who wants the other 3.4 per cent does not need Mistral's permission, or its guardrails, which are a property of the hosted API rather than the file on your disk. Open weights and safety filters are structurally separable, and every lab shipping open weights knows it. Mistral said so out loud on launch day, which is more than most.

Three reasoning modes

Low, Medium, and High - High being the default. Low replaces the old "None" mode. This is the same performance-per-dollar dial every lab is now building, and it exists because inference cost is the thing enterprise buyers actually argue about.

What this is really for

"Trained on European infrastructure, clear lineage from inception to release" is not a technical claim. It is a procurement argument. The buyers Mistral is courting are governments and regulated industries that cannot send data to a US API and cannot explain to an auditor what happened inside a model they cannot inspect.

A 1T-parameter model you can self-host at open-weight prices, scoring at parity with the best Chinese open models on coding, is the strongest product that argument has had. Key Takeaway: the headline is the parameter count; the substance is the price-per-token on a model you are allowed to own. The caveat is written into the launch post - the guardrails are still being tuned, and once the weights are downloaded, tuning them is no longer Mistral's decision.