Mistral Small 3

Mistral Small 4, the lab's hybrid model unifying instruct, reasoning and coding in one efficient set of weights. It is the first Mistral to fold the Magistral reasoning, Pixtral multimodal and Devstral agentic-coding lines into a single versatile model, tuned for general chat, coding, agentic tasks and complex reasoning. It is a Mixture-of-Experts with 119B total and roughly 6B active parameters, taking text and image input across a 256K context.

Small 4 is the release where Mistral stopped asking people to pick a model. The lab describes it as the first Mistral to fold its reasoning, multimodal and agentic-coding lines into one set of weights, so a fast instruct reply, a step-by-step derivation and a tool-driven coding run all come from the same checkpoint. Its sparse design routes four of 128 experts per token, which is why a 119 billion parameter model costs what a much smaller one does to serve.

The control that makes the merge work is a reasoning-effort parameter with two named settings, and Mistral describes both in terms of the models they replace. Setting it to none produces fast, lightweight answers in the same chat style as the previous Small release. Setting it to high produces deep step-by-step reasoning with roughly the verbosity of the lab's dedicated reasoning models. That is the whole switch: no second endpoint, no second deployment.

Mistral's own comparison is against its earlier models rather than the field. With reasoning on it reports 71.2 on GPQA Diamond and 78 on MMLU Pro, against 59.1 and 73.5 for the same weights in instruct mode, 48 against 35.7 on the AllenAI instruction-following benchmark IFBench, and 60 against 46.3 on the vision benchmark MMMU-Pro. Alongside the accuracy the lab pushes an efficiency claim: it says Small 4 matches or beats a comparable open model on three benchmarks while generating substantially shorter answers, and that shorter answers are the point, because they are what actually shows up as latency and cost.

The engineering numbers are stated plainly. Mistral reports a 40% cut in end-to-end completion time in a latency-tuned setup and three times the requests per second in a throughput-tuned one, against the previous Small generation. It names the minimum hardware - four H100s, two H200s or a single DGX B200 - and ships under Apache 2.0, with day-one support across the common open serving stacks. The lab's own framing of who it is for is equally plain: developers doing coding automation and codebase exploration, enterprises running chat assistants and document understanding, researchers doing maths and complex reasoning. It is the efficient tier, not the frontier one, and Mistral points at its larger models for the hardest work.