Mistral Large 4

Overview. Mistral Large 4 is Mistral AI's frontier multimodal model, built on a granular Mixture-of-Experts architecture containing 1.05 trillion total parameters and an integrated 1.6B vision encoder. Running with 52 billion active parameters per token, it supports combined text and visual inputs across a 512K-token context window with up to 262K output tokens. Developers reach for it when demanding frontier analytical capability and visual comprehension without sacrificing European data sovereignty.

Key strengths. The model demonstrates high analytical aptitude across mathematics, research, and code synthesis, reinforced by configurable chain-of-thought reasoning that can be toggled to high effort. Its 52B active parameter execution profile and native context caching deliver responsive interactive generation across massive prompts. Multimodal image understanding allows developers to inspect rich visual data and technical diagrams directly within complex reasoning pipelines.

When to use it. Route to Mistral Large 4 for multi-step mathematical derivations, extensive codebase reviews, and document research requiring combined visual and textual processing. It fits interactive reasoning workflows where deep chain-of-thought traces are needed on demand without slowing conversational execution. Use it within the Smart Router whenever workloads require frontier reasoning under strict European compliance standards.

Limitations. Mistral AI has not published the training cutoff date for this release, and the model does not support audio inputs or outputs. Furthermore, while technical documentation mentions a 1M-token ceiling, current API endpoints cap generation at a 512K context window.

Sovereignty note. Mistral Large 4 is made by Mistral AI, a French company under EU jurisdiction, so the model's creator is not subject to the US CLOUD Act. Whether a request stays sovereign depends on the Inference Provider and region that serve it: check the Provider Endpoints below before you send sensitive data.