Mistral Small 4 Unifies Text, Vision, and Reasoning in Single 119B Model
Mistral AI launches Mistral Small 4, a 119-billion-parameter Mixture-of-Experts model that merges instruction, reasoning, coding, and multimodal capabilities into one system. The model features a 256k context window and adjustable reasoning effort to balance speed and accuracy.
Mistral AI releases Mistral Small 4, a new 119-billion-parameter Mixture-of-Experts model that combines multiple capabilities into a single deployment. The model replaces the need for separate specialized systems by merging the instruction-following of Mistral Small, the reasoning of Magistral, the vision capabilities of Pixtral, and the coding skills of Devstral into one unified architecture.
The model utilizes 128 experts with only 4 active experts per token, resulting in 6 billion active parameters per token for efficient processing. It supports a massive 256k context window, which simplifies enterprise workflows by reducing the need for complex chunking and retrieval strategies during long-document analysis, codebase exploration, and multi-file reasoning tasks.
A standout feature of Mistral Small 4 is its configurable reasoning effort, allowing developers to adjust test-time computation per request. Users can set the reasoning effort to none for fast, chat-style responses, or increase it to high for deliberate, step-by-step problem solving, providing flexible control over latency and output depth.