Mistral Small 4

Mistral Small 4 is the first release in the Mistral Small line to fold reasoning, multimodal, and coding abilities into a single hybrid model, removing the need to switch between separate instruct, reasoning, and vision assistants. The architecture is a Mixture-of-Experts design with 128 experts activating 4 per token, totaling 119B parameters with roughly 6B active per forward pass (about 8B when embedding and output layers are included). That sparse activation pattern lets the model handle long, complex workflows without the full parameter cost being incurred on every token, and it is released openly under Apache 2.0, continuing Mistral's tradition of openly licensed weights and customization.

Practically, the model targets general chat, agentic coding, document understanding with images, and step-by-step reasoning, with configurable reasoning effort that lets teams trade speed for depth depending on the task. It sits inside the cataloged API limit context window and was developed as a founding member of the NVIDIA Nemotron Coalition, giving it day-one availability across Mistral's own API, open-weight channels, and NVIDIA NIM. Independent scoring places its intelligence index at 19.6, positioning it as a versatile open alternative for teams that want one model covering assistant, reasoning, and multimodal roles rather than stitching together separate specialized deployments.