DownloadMistral Medium 3.5 128B is Mistral AI’s first flagship merged model, unifying instruction following, reasoning, coding, vision, and agentic capabilities within a single set of weights. It uses a dense 128B-parameter architecture and supports a 256K-token context window for large documents, codebases, and extended workflows. The model accepts both text and images while generating text, using a vision encoder trained from scratch to accommodate variable image sizes and aspect ratios. Reasoning effort can be configured per request, allowing users to switch between fast direct responses and higher-compute reasoning for complex problems and agents. It includes native function calling, structured JSON output, strong system-prompt adherence, and multilingual support across dozens of languages. Mistral positions it as the successor to Medium 3.1, Magistral, and Devstral 2, consolidating their specialized capabilities into one general-purpose model.
Features
- 256K-token context window
- Dense 128B-parameter architecture
- Native text and image input with text output
- Configurable instant and high-reasoning modes
- Native function calling and structured JSON output
- Advanced agentic coding and tool-use capabilities
- Multilingual support across dozens of languages
- Compatible with vLLM, SGLang, Transformers, and llama.cpp
Project Samples
