Mistral Medium 3.5

Mistral Medium 3.5 is the model the French lab actually pushes for day-to-day work, and it is a deliberate departure from the industry's habit of shipping a separate model for every job. Instead of a chat model, a reasoning model and a coding model, Medium 3.5 is one dense 128-billion-parameter network in which the reasoning effort is a parameter the caller sets. It handles text and images and works with a 256,000-token context window. The model was released on 29 April 2026 (versioned v26.04 in Mistral's catalogue) and replaced Mistral Medium 3.1, which Mistral deprecated on 22 May 2026 and switches off on 31 August 2026. Where it stands out is agentic and software work: reported figures put it at 77.6% on SWE-Bench Verified — within about a point of the best proprietary models of the period — and 91.4% on the Tau-3 Telecom benchmark for multi-step tool calling. API pricing is 1.50 US dollars per million input tokens and 7.50 per million output, several times the rate of the older Medium 3 but well below frontier proprietary tiers. What separates it from most models in that performance band is the licence. Weights are published under a modified MIT licence, so an organisation that cannot send user content to a third-party API can run the same model inside its own infrastructure. That combination — near-frontier coding scores, a European vendor and self-hostable weights — is the argument Mistral makes to regulated European buyers, and it is a different argument from the one American labs make. Access is available through the Mistral API, the Le Chat assistant and Hugging Face.

▸Videos

No videos yet.

▸Reviews