Mistral Large 4 (Le Chonk) releases its weights: a trillion parameters and three modes

⚡ The gist in 5 seconds

• Mistral has released the weights of Mistral Large 4 (Le Chonk): a trillion parameters, natively multimodal.
• In coding and defensive cybersecurity, the model has taken the lead among open models — FrontierCode, Terminal-Bench 4.0.
• Three reasoning modes: Low, Medium, and High, with High on by default.

🔍 What happened

Mistral has released the weights of Mistral Large 4 — internally the company calls the model Le Chonk. It's a trillion-parameter flagship, natively multimodal. It had been sitting in public preview since the beginning of the month, and now the weights are out — post on Mistral's blog. Three weeks of extra RL passed between the preview and the release.

Here are the numbers from the post. FrontierCode 2.0 by Cognition — a private benchmark that checks whether a maintainer would merge the agent's patch — gives 19.7%. Artificial Analysis Coding Agent Index — 48%, exactly on par with GLM-5.3, but with 47% fewer tokens per task. Terminal-Bench 4.0 — 41%, up 13 percentage points from the preview and 30% fewer output tokens. In blind pairwise comparisons on Surge Code, the model beat GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, losing only to Claude Opus 5.5.

The mechanics are simple. There are now three reasoning modes: Low for simple tasks (it replaced the old "None"), Medium for everyday use, and High as the default — for the toughest cases. The model was trained from scratch on European infrastructure, and you can deploy it on your own, without Mistral's cloud.

🧠 Our take

The bet on "sovereign AI" here isn't a marketing line — it's the direct reason for the release. Open weights plus European training is an argument for banks, government agencies, and defense contractors that can't route their data through someone else's API. Our assessment: Mistral isn't selling a benchmark — it's selling usage rights.

The second layer is token economics. 47% fewer tokens at the same index score matters more than a "we're number one" line. An agent that burns half the context per task is cheaper in production, and you see that in your monthly bill, not on a leaderboard. We believe this is the number that will matter most in conversations with customers.

Who wins: teams that need a strong coding model inside their own perimeter, and European integrators. Who loses: vendors selling model access only via API — they'll have some explaining to do on price. Also losing are those chasing the top spot on the leaderboard: the gap with GLM-5.3 here is zero, but the token difference isn't.

⚠️ What we don't know

The post contains no license, no hardware requirements, and no inference pricing. A trillion parameters — what does that mean for running it yourself? Not a word. FrontierCode 2.0 is a private benchmark that can't be verified from the outside. Comparisons are only against the models invited to the measurement. And there's a single source — Mistral's own blog, i.e., an interested party.

Prediction: within a quarter, there will be at least one independent replication of the Terminal-Bench 4.0 and Coding Agent Index measurements — provided the weights are actually downloadable without a corporate gate. If the numbers match up, API prices for open coding models will dip.

Source: mistral.ai