Mistral Small 4

Mistral Small 4, released on 16 March 2026, is where the French lab stopped maintaining three separate small models. Magistral did reasoning, Devstral drove coding agents and Mistral Small handled ordinary instructions; Small 4 replaces all three. It is a sparse mixture-of-experts network with 119 billion parameters in total but only 6 billion activated per token (8 billion counting embedding and output layers), routed through 128 experts with four active at a time, and it takes both text and images with a 256,000-token context window. The sparsity is the point. Against GPT-OSS 120B — the obvious open-weight comparison — Mistral reports matching or better results on AA LCR, LiveCodeBench and AIME 2025, while producing markedly shorter answers: 0.72 on AA LCR from 1,600 characters of output, where comparable Qwen models need close to four times as much text to land in the same place. Shorter outputs mean lower latency and lower cost per answered question, and against Mistral Small 3 the lab claims 40% lower latency and triple the throughput. Hosted pricing is 0.15 US dollars per million input tokens and 0.60 per million output. Weights are published under a plain Apache 2.0 licence, with no field-of-use or regional clauses, which puts the model in the small group of genuinely unrestricted releases at this capability level. Self-hosting is not trivial — Mistral's own guidance starts at four NVIDIA HGX H100 systems, two HGX H200s or a single DGX B200 — so the practical audience is organisations that already run GPU infrastructure and cannot route user content through someone else's API.

▸Videos

No videos yet.

▸Reviews