Mistral said on 16 March 2026 that it had released Mistral Small 4, the next major model in its Small family, under the Apache 2.0 licence.

Small 4 is the first Mistral model to unify Magistral-style reasoning, Pixtral multimodal input and Devstral agentic coding in one system. It is a mixture-of-experts model with 119 billion total parameters, 6 billion active per token, 128 experts with four active, a 256k context window and native text-plus-image inputs.

Users can set reasoning_effort to “none” for fast chat-style replies or “high” for deeper step-by-step reasoning. Mistral reported a 40% cut in end-to-end completion time in a latency-optimised setup and 3× more requests per second versus Mistral Small 3 in a throughput-optimised setup.

Minimum infrastructure cited includes 4× NVIDIA HGX H100, 2× HGX H200 or 1× DGX B200. The model is available via Mistral’s API and AI Studio, Hugging Face, build.nvidia.com, and as an NVIDIA NIM, with community support on vLLM, llama.cpp, SGLang and Transformers.

Mistral said it joined the NVIDIA Nemotron Coalition as a founding member alongside the release, and that with reasoning enabled Small 4 matched or beat GPT-OSS 120B on several benchmarks while producing shorter outputs.


Share this post
The link has been copied!