mistral-medium-latest

Overview. Mistral Medium 3.5 (the 2604 snapshot) is a dense 128B-parameter model from Mistral AI. It accepts text and image input and produces text output. Reasoning can be switched on or off per request, so one model covers both quick instruction-following and longer reasoning-heavy work.

Key strengths. The model has a 262,144-token context window, confirmed in Mistral's published configuration, which suits long documents and extended multi-turn sessions. Because it is dense, all 128B parameters are active on every token. Native image input means screenshots, diagrams and scanned pages can go in the same request as text. We have no benchmark results for this snapshot, so these strengths come from its architecture and documented capabilities, not from scores.

When to use it. Use it for complex tasks that need a large context window, such as analysing long documents or working across large codebases. It also fits workloads that mix text and images. Turn reasoning on for harder problems and off when you want a faster, more direct reply. If a task is simple and latency-sensitive, a smaller model routed through the Smart Router is likely the better choice.

Limitations. No benchmark data, training-data cutoff or separate maximum output length has been established for this snapshot, so validate quality on your own workload before committing. Reasoning is a simple on/off toggle rather than graded effort levels. The Modified MIT licence is not standard MIT: it restricts use above a company-revenue threshold unless you obtain a separate commercial licence from Mistral AI.

Sovereignty note. Mistral AI is headquartered in France and carries no CLOUD Act exposure, which supports Sovereign EU deployments. Endpoint data residency and training-on-input policy are not confirmed in the research for this model, so check them on the specific endpoint you use.