Mistral AI has launched Mistral Large 4 in public preview. The natively multimodal Mixture-of-Experts model has 1.05 trillion total parameters, activates 49 billion per token, and supports a one-million-token context window.
Mistral positions ML4 as its strongest model yet for coding, agentic workflows, visual understanding, cybersecurity, finance, and legal work. Developers can access it through Mistral Studio and the API, with support for function calling, structured outputs, built-in tools, and Mistral’s Agents and Conversations endpoints.
The model is currently API-only. Mistral says it will release the weights, additional architectural details, benchmarks, and post-training information by the end of October. Its current performance comparisons remain vendor-reported and should be treated as preliminary until independent evaluations and the promised weights arrive.
Why it matters
Mistral Large 4 could give builders another serious open-weight foundation for multimodal agents and specialized enterprise systems. Its sparse architecture is especially notable: only a fraction of the trillion-parameter model is active for each token.
The planned weight release also matters for organizations seeking more control over deployment, infrastructure, and data location. It extends the direction Mistral established with Mistral Medium 3.5’s cloud and coding-agent push into a much larger multimodal foundation.
For now, developers should evaluate the preview on their own workloads rather than treating launch benchmarks as settled evidence. API behavior can be tested today, but deployment assumptions should wait for the promised weights, technical details, and independent results.