Mistral Medium 3.5

Mistral Medium 3.5 is Mistral AI's flagship foundation model, launched on 29 April 2026. It is a dense (non-mixture-of-experts) transformer with 128 billion parameters and a 256,000-token context window, handling text and image input across instruction-following, reasoning, and agentic coding in a single model. Key benchmarks include a 77.6% SWE-bench Verified score, beating all closed-source frontier models except Claude Sonnet 4.6 and DeepSeek V4 Pro, while running on as few as four GPUs for self-hosted deployments. API pricing is $1.50 input / $7.50 output per million tokens.

The model's architecture strategy consolidates three previously separate Mistral lines: Medium 3.1 (general tasks), Magistral (reasoning), and Devstral 2 (agentic coding) are replaced by a single set of weights with configurable reasoning effort per request. It is released as an open-weight model, meaning weights are available for self-hosting, a significant competitive differentiator versus closed-source US counterparts for European organisations with on-premises deployment requirements.

Mistral Medium 3.5 is the technical engine behind Le Chat Enterprise's data-residency-compliant deployments. Its availability as an open-weight model makes it deployable in fully air-gapped environments — critical for European defence, healthcare, and intelligence use cases where cloud connectivity is prohibited. Combined with its coding capability score, it positions Mistral as a credible alternative to Anthropic and OpenAI for EU public procurement.