Mistral AI has released Mistral Small 4, a new open-source artificial intelligence model that combines reasoning, multimodal processing and coding capabilities in a single system. The company reports that users no longer need to switch between separate specialised models for different tasks.
The model uses a Mixture of Experts architecture with 128 specialists, activating only four per processed unit of text. This approach allows the system to scale efficiently while keeping resource use low. The total parameter count stands at 119 billion, with 6 billion active at any given moment.
A key feature is the configurable reasoning effort. Users can set the model to deliver quick, lightweight answers for everyday tasks or switch to a deeper, step-by-step reasoning mode for complex problems. Mistral claims a 40 percent reduction in response time and three times more requests per second compared to its predecessor.
The model accepts both text and image inputs and supports a 256,000-token context window, making it suitable for long documents and visual analysis. Minimum hardware requirements include four NVIDIA HGX H100 units.
Mistral Small 4 is available under the Apache 2.0 licence, meaning it can be freely used, modified and distributed. It is accessible through the Mistral API, Hugging Face and NVIDIA’s build platform.
