Mistral Small 4 unifies reasoning, image input and agent coding

Mistral released Small 4 on 16 March, combining capabilities previously spread across its reasoning, vision and coding models. The Apache 2.0 mixture-of-experts model has 119 billion total parameters and activates six billion per token, excluding embedding and output layers. It supports a 256K context window and both text and image input.

A reasoning-effort setting lets applications switch between faster responses and more involved work. The company offers API access and downloadable weights, with serving support across several open inference tools. The Small name should not be read as a promise that the full model fits a typical phone or laptop: total weight storage and infrastructure still matter even when only part of a mixture-of-experts network is active during inference.