Mistral Small 4

One model instead of three

Mistral Small 4 was released on 16 March 2026, and its main organising idea is consolidation: it unifies the company's reasoning line (Magistral), its multimodal line (Pixtral) and its agentic coding line (Devstral) in a single model^[1]. Mistral describes it as a hybrid model that brings instruct, reasoning and coding together in one efficient package^[1]. The practical consequence of that design is that a team no longer has to stand up separate deployments for chat, for hard problems that need deliberation, and for tool-using coding work; one model covers all three, which simplifies both the infrastructure and the choice of which checkpoint to serve.

Architecture

Underneath, Small 4 is a mixture-of-experts (MoE) model with 128 experts, of which 4 are active for each token^[1]. It has 119B parameters in total and 6B active parameters, a figure that rises to 8B once the embedding and output layers are counted^[1].

The MoE idea is worth spelling out, because the parameter counts are easy to misread. A dense model of 119B parameters does arithmetic with all of them for every token it generates. A sparse mixture-of-experts model keeps the same large pool of weights but routes each token through only a small subset of specialised sub-networks, so the compute cost per token tracks the active count while the knowledge capacity tracks the total. That is why a model can carry 119B parameters and still be cheap enough to serve at high volume.

The context window is 256K tokens, and the model accepts both text and image input^[1]. Long context and multimodality are what make the Pixtral heritage usable in practice: a single request can carry a long document or a screenshot alongside the question.

The reasoning switch

Rather than shipping separate "instruct" and "reasoning" checkpoints, Small 4 exposes a `reasoning_effort` parameter^[1]. Set to "none", the model answers directly and quickly; set to "high", it works through the problem step by step before answering^[1]. This turns reasoning from a property of the model into a per-request setting: routine lookups can take the fast path, while maths, multi-step analysis or tricky code can pay for extra deliberation. The same switch is what lets the model serve as a coding agent, where planning ahead usually beats answering immediately.

Speed, price and hardware

Mistral positions the model on efficiency as much as capability. Compared with Mistral Small 3, it offers up to 40% lower latency and handles 3x more requests per second^[1].

The minimum hardware to run it is 4x H100, 2x H200 or 1x DGX B200^[1] — a reminder that an open-weight model with 119B total parameters still needs a serious multi-GPU machine even when the per-token compute is small. Through the API it is priced at $0.15 per million input tokens and $0.60 per million output tokens^[1]. The weights are released under Apache 2.0 on Hugging Face and as an NVIDIA NIM^[1], so the same model can be used as a hosted endpoint, a downloadable checkpoint, or a packaged NVIDIA deployment.

Reported results

Mistral says the model matches or exceeds GPT-OSS 120B on three benchmarks while producing shorter outputs, and that it scores 0.72 on AA LCR^[1]. The shorter-output detail matters for the economics: output tokens are the expensive and slow part of inference, so reaching a given quality with fewer generated tokens improves both latency and cost.

Where it sits

Mistral's documentation lists Small 4 as the replacement for Mistral Small 3.x, Magistral Small and Mixtral^[2]. That is a notable amount of retired ground: an instruct model, a reasoning model and the older Mixtral generation all hand their role to one entry. Alongside the release, Mistral said it joined the NVIDIA Nemotron Coalition as a founding member^[1]. Taken together, the release is a statement about how the company now builds models — fewer, broader checkpoints with behaviour controlled by inference-time settings rather than by separate releases for each capability.