Mistral Large 4: 1T parameters, open weights by end of October

Europe's underdog just shipped something big. Mistral released Mistral Large 4 today, nicknamed "Le Chonk", and for once the nickname undersells it: one trillion parameters, trained from scratch in Mistral's own European data centers.

What Mistral announced

• Size: 1 trillion parameters in a mixture of experts, 49 billion active per token. Mistral calls it a hybrid instruct and reasoning model.
• Input: natively multimodal, text and images in, text out.
• Languages: more than 160, every official EU language among them.
• Where: a public preview API on Mistral Studio, today. Mistral lists $1.36 per million input tokens and $4.18 per million output tokens.
• Weights: public at the end of October. VentureBeat reports October 27, under a custom Mistral license.
• Training: about two months on Nvidia Grace Blackwell GPUs. Mistral's post says 3,800 of them, some coverage rounds that to 4,000.

The numbers that matter if you code

All of these are Mistral's own figures, not independent tests yet:

• DeepSWE v1.1: 61.7%
• SWE-Atlas-QnA: 59.4%
• Terminal Bench 4.0: 28.3%
• Cybench: 93% of 40 security challenges

The pitch is security work. Mistral argues that defending software often starts with proving a flaw is real, "exactly the kind of work safety filters in closed models can block." The claim on top: the best open weights model from the US or Europe on aggregated benchmarks.

Our take

A trillion-parameter model you can download is a real event, and a European one is rarer still. But read the coding numbers with care: every one of them comes from the company selling the model, and the headline claim is about open models, not the closed ones at the top. The interesting part arrives with the weights: when anyone can run and measure it, we will know where it really stands.

In AyCode

AyCode already runs Mistral Vibe, Mistral's own coding agent, and OpenCode, which can call Mistral's API with your own key. Run them side by side with Claude Code or Codex and give all of them the same task. That tells you more about a new model than any chart.