Mistral Medium 3.5 Released

Mistral Medium 3.5 is the first flagship merged model, featuring 128B dense parameters and a 256k context.

• It unifies instruction following, reasoning, and coding into a single set of weights, and reasoning_effort can be adjusted per request.
• It's a multimodal architecture that accepts text and image inputs, supports multiple languages including Korean, and features a newly trained vision encoder to handle various image sizes and aspect ratios.
• In Le Chat, it replaces Mistral Medium 3.1 and Magistral, and in Vibe, it replaces Devstral 2.
• On public benchmarks, it posted τ³-Telecom 91.4% and SWE-Bench Verified 77.6%, and is distributed under a Modified MIT License with exceptions for high-revenue companies.
• vLLM nightly is recommended for local inference, and an EAGLE model is also provided to boost speed.