GitHub radar

Mistral Medium 3.5: reasoning + coding + vision in one

Mistral AI's first flagship merged model: instruction-following, reasoning, and coding in a single open-weight 128B — plus vision and 256K context.

01mistralai/Mistral-Medium-3.5-128B 438139k downloads/mo128B paramsimage-text-to-text

Mistral Medium 3.5 is Mistral AI's first 'merged' flagship: a 128-billion-parameter dense model that combines instruction-following, multi-step reasoning, and code generation in a single set of open weights. Previous Mistral models were specialized — Devstral for code, Mistral Large for dialogue. Medium 3.5 does all three in one request, plus it understands images via a custom-trained vision encoder. Context window is 256,000 tokens. On SWE-Bench Verified (real GitHub bug-fixing tasks) it scores 77.6%; on the agentic τ³-Telecom benchmark — 91.4%. Reasoning effort is configurable: it can switch off for fast replies and engage fully for complex problems. Weights are open under a modified MIT license; self-hosting requires approximately four GPUs.

Why a vibe-coder should care

Instead of picking the right Mistral model per task, one model now covers all three: images, reasoning, and code — in a single request. Available via the Mistral API or for free through Le Chat's Work mode for multi-step tasks.

How to install

Copy this and send it to your agent — Claude Code, Codex, any of them:

Connect me to Mistral Medium 3.5 by Mistral AI: https://huggingface.co/mistralai/Mistral-Medium-3.5-128B — set up access via the Mistral API (api.mistral.ai), show me how to send a request with an image and a coding task.

Серверная модель — дома не запустить: через Ollama занимает около 80 ГБ памяти. Для большинства удобнее через Le Chat или API Mistral.

Open on Hugging Face