mistral-medium-3-5

Mistral Medium 3.5 is our first "fully merged" model with enhanced code and agentic capabilities. It demonstrates strong tool dexterity, making it suitable for both VIP (Vibe Code Scaffold) and agentic research tasks.

Mistral Medium 3.5 is a 128B-parameter Dense model that unifies instruction-following, reasoning, and coding capabilities with multimodal input support, 256k context, and native function calling under Modified MIT license.

Mistral Medium 3.5 is a powerful hybrid model that unifies the capabilities of three model families — Instruct, Reasoning (previously Magistral), and Devstral — into a single model. Built on a Dense architecture, it delivers 128B total parameters enabling strong performance with efficient inference. The model supports multimodal input (text and images), a 256k token context window, and native function calling with JSON output. Users can switch between a fast instant-reply mode and a reasoning mode with configurable reasoning effort, allowing test-time compute scaling when needed.

It supports 40+ languages including English, French, German, Spanish, Italian, Portuguese, Dutch, Chinese, Japanese, Polish, Arabic, Farsi, Urdu, Hebrew, Turkish, Indonesian, Lao, Malay, Thai, Tagalog, Vietnamees, Hindi, Bengali, Gurjarati, Kannada, Marathi, Nepali, Punjabi, Tamil, Telugu, Breton, Catalan, Czech, Danish, Greek, Finnish, Croation, Norwegian, Romanian, Swedish, Serbian, Ukranian, Korean and Russian. The vision encoder is trained from scratch to natively support variable image sizes and aspect ratios using RoPE-2D rotary position encoding. Mistral Medium 3.5 is ideal for general chat assistants, coding agents, document parsing and extraction, research assistants, and fine-tuning for specialized tasks.