Mistral Large 4 preview: a trillion-parameter open-weight model, weights still to come
Mistral has previewed its biggest model yet, Mistral Large 4, with a 1 million token context window. It's meant to be open-weight, but the weights aren't public yet.
By ScoviAI Team
Key points
- Mistral Large 4 entered public preview on October 6, 2026
- About 1 trillion total parameters, with around 50 billion active at a time
- 1 million token context window; it can read images too
- The weights aren't public yet; Mistral says more details will come first
Mistral, Europe's best-known AI company, has released a public preview of Mistral Large 4, its largest model so far. Mistral describes it as an open-weight, general-purpose model that can work with both text and images, with a 1 million token context window.
The headline promise is "open-weight": once released, anyone will be able to download the model and run it on their own computers. That hasn't happened yet.
How big is Mistral Large 4?
Mistral Large 4 uses a Mixture-of-Experts design. The model is huge, but only part of it works on each word, which keeps it faster and cheaper to run.
| Mistral Large 4 | |
|---|---|
| Total parameters | About 1 trillion (Mistral's docs list 1.05 trillion) |
| Active parameters | About 50 billion (52 billion in the docs, 49 billion on the research page) |
| Context window | 1 million tokens |
| Inputs | Text and images |
Mistral's own pages give slightly different numbers, as the table shows.
How good is it?
In a blind human evaluation of coding answers run by Surge AI, the preview ranked second of five models, behind Anthropic's Claude Opus 5 and ahead of Kimi K3 and two GLM models. That's one test on one kind of task, so it's an early signal rather than a verdict.
Made in Europe
Mistral says it trained the model on its own infrastructure in Europe, and the preview runs there too. For European companies and governments that want AI kept under European control, that's a big part of the appeal.
When do the weights come out?
Mistral hasn't given a firm date. It says it will first share more about the model's architecture, more benchmarks and how it was trained, as it works toward releasing the weights. If you're wondering why open weights matter, read Open-weight AI models, explained.
What it means for you
- Developers can try the preview through Mistral's platform now.
- Companies that need to run AI on their own servers should watch for the weights release, which would make this one of the largest models anyone can host themselves.
Sources
- Introducing Mistral Large 4 (Mistral AI)
- Mistral Large 4 model card (Mistral docs)
ScoviAI Team is the ScoviAI newsroom. We write every article ourselves from the original announcements and link our sources. Spotted a mistake? Tell us and we'll fix it.
ADVERTISEMENT
