Mistral AI Introduces Mistral-Medium-3.5-128B As One Unified Tool
Mistral-Medium-3.5-128B is a dense flagship model designed to handle complex reasoning, coding, and instruction-following tasks. It serves as a unified replacement for several previous models released by the company.
The development team at Mistral AI created this version to provide better performance across multiple specialized functions within a single set of weights. This approach aims to simplify workflows by offering one versatile tool instead of multiple separate models.
Unified reasoning and multimodal capabilities
- Dense architecture with 128 billion parameters.
- Large 256k token context window for long documents.
- Multimodal support for processing both text and images.
- Configurable reasoning effort per individual request.
- Native function calling and JSON output capabilities.
- Support for dozens of different languages.
Developers building autonomous agents or complex coding tools can utilize the adjustable reasoning settings to balance speed and depth. Those working with large datasets or long-form documents will benefit from the expansive context window that allows for much more information to be processed at once.
Technical notes and performance fixes
The team noted that an earlier version of the configuration caused issues with how the model performed during long-context sessions. To ensure the best results, users should use the updated Transformers configuration and avoid older GGUF files generated before the fix.
For those running the model locally, using vLLM or SGLang is recommended to increase speed.
"Mistral Medium 3.5 is our first flagship merged model,"
You can explore the capabilities of this new model by visiting its page on Hugging Face.
