Mistral AI Releases Mistral Small 4: A Versatile Model with Reasoning and Multimodality

Mistral AI has announced Mistral Small 4, a new model in the Small family that, for the first time, combines the capabilities of the company's flagship models: Magistral (reasoning), Pixtral (multimodality), and Devstral (agentic coding). Users can adjust the level of reasoning effort via the `reasoning_effort` parameter, choosing between fast responses and deep multi-step reasoning. The model uses a Mixture-of-Experts architecture with 128 experts (4 active per token), totaling 119 billion parameters, of which 8 billion are active (including embedding and output layers). The context window is 256k tokens. Compared to Mistral Small 3, end-to-end latency is reduced by 40%, and throughput is increased by 3 times. Mistral Small 4 outperforms GPT-OSS 120B in efficiency, generating significantly shorter responses with comparable or better quality. The model is released under the Apache 2.0 license and is available via the Mistral API, AI Studio, and as an NIM on the NVIDIA build.nvidia.com platform. Mistral Small 4 supports text and image inputs, enabling its use for chat, coding, document analysis, and complex reasoning. The company also announced that it has joined the NVIDIA Nemotron Coalition as a co-founder.