← All posts
This post is written in English.
Mistral Large 4, announced on 6 October, is a 1-trillion-parameter multimodal model with 52 billion active at a time. Mistral says it trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacentres.
The numbers
Mistral's public preview API costs $0.70 per million input tokens and $2.10 per million output, with cached input at $0.07. The post promised open weights "by the end of the month". It doesn't name a licence, and I haven't confirmed the release, so check before planning around it.
| Benchmark | Reported score |
|---|---|
| DeepSWE v1.1 | 61.7% |
| Terminal-Bench 4 | 28.3% |
| Cybench | 93% (self-reported) |
| Surge AI blind coding eval | 3.74/5, second of five |
Most headline numbers come from Mistral itself. Others are attributed to Artificial Analysis, vals.ai, Surge AI and Lakera, but the post doesn't say how independently those were run.
Why "sovereign" matters here
The pitch isn't only about scores. Mistral says it can run a European deployment end to end from its own datacentres. That matters to governments and regulated companies in ways a benchmark can't capture, especially with EU AI Act deadlines approaching.
My read
I'm more interested in the infrastructure than the leaderboard. Owning the GPUs and the datacentres is a much bigger commitment than releasing a model, and it's what makes the sovereignty claim more than marketing. Whether it's the right model for a given job is something only your own evaluation can tell you.
