Mistral Large 4

Verified

Mistral's trillion-parameter multimodal open-weight model, in public preview

Visit Mistral Large 4 →No rating yetPaid

Product preview

Mistral Large 4

What it is

Mistral Large 4 is a large language model from French company Mistral AI, released as a public preview on October 6, 2026. Mistral describes it as a natively multimodal mixture of experts model with about 1 trillion total parameters and 52 billion active, trained from scratch in its own European datacenters. It targets agentic coding, long business workflows, image and document understanding and cybersecurity tasks, and it was trained on data spanning more than 160 languages. The preview is served through Mistral Studio and the Mistral API with usage based pricing, and Mistral has said the open weights will follow for self-hosted and on-premise deployment.

Key features

Mixture of experts with 52B active parameters
Native image and document understanding
Agentic coding and tool use
Planned open weights for self-hosting

What we liked

+ Strong focus on open weight deployment
+ European infrastructure and data residency
+ Broad multilingual training

Worth noting

· Still a preview, so behavior may change
· Weights and license not yet published at launch

Editor's notes

More context from AI BirdView

Mistral Large 4 is notable as Mistral's biggest model so far and as a serious open weight option built and served in Europe. The company frames it around control: organizations should be able to run a frontier class model on their own infrastructure, under their own policies, rather than depending on a closed provider.

Mistral reports competitive results on coding benchmarks such as DeepSWE and Terminal Bench, and says it leads open weight models built in the US or Europe.
The model combines visual grounding with agent behavior, for tasks like inspecting technical drawings or large satellite images.
Mistral positions it strongly for security work, saying it can reproduce and patch real vulnerabilities in testing.
Training data covered more than 160 languages, including every official EU language.
It shares the training and reinforcement learning setup that Mistral offers enterprises through Mistral Forge.

Mistral Large 4 is best for developers and enterprises that want a capable general model with a path to self-hosting, especially in regulated European settings. A practical tip: test it through the preview API on your own evaluation set now, and plan a second round once the weights and final post-training land, since Mistral says the model is still improving.