Mistral Large 4 Enters Public Preview with a Trillion-Parameter MoE Design

Mistral AI launched a public preview of Mistral Large 4 on October 6, 2026. The French company's new multimodal model targets coding, agentic workflows and enterprise tasks.

The model uses a mixture-of-experts architecture with roughly 1.05 trillion total parameters. About 49 billion parameters are active for an inference step according to the launch announcement. Total parameter count should not be confused with compute used per token.

Mistral Large 4 handles multimodal inputs, including text and images, for applications such as document understanding and technical analysis. Specific supported formats should be checked in current model documentation.

Documentation lists a context window of up to one million tokens. Large context capacity is useful but does not guarantee reliable retrieval or reasoning over every long document.

The API is available in public preview through Mistral Studio. The company said model weights would be released toward the end of October; the preview and planned open-weight distribution are distinct milestones.

Enterprises should compare hosted API use with self-managed inference, taking into account GPU requirements, latency, security, auditing and licensing.

Vendor benchmarks are useful reference points, but organizations need evaluations on their own tasks and data, including failure rates and costs.

The upcoming weight release and deployment ecosystem will be important tests of how competitive a European open-weight frontier model can become.