What it is
Mistral Small 4 (API ID mistral-small-2603) was released on 16 March 2026 and is still Mistral's current Small model; it replaced Small 3.2 and Magistral Small 1.2, both retired on 31 July 2026. It is a 119-billion-parameter mixture-of-experts model with 128 experts, four active per token, so about 6 billion parameters do the work for each token. It merges Mistral's earlier reasoning (Magistral), vision (Pixtral) and coding (Devstral) lines, reads text and images, writes text and has a 256K-token context.
A reasoning_effort setting switches between quick replies (none) and step-by-step reasoning (high). The weights are free under Apache 2.0, with an FP8 checkpoint, an NVFP4 version and community GGUF builds; Mistral lists four H100s, two H200s or one DGX B200 as the minimum to serve it.
The API costs $0.15 per million input tokens and $0.60 per million output tokens, with 50% off through the batch API and up to 90% off repeated input.
Our review
Small 4 is a price-and-licence choice more than a capability one. Artificial Analysis scores the reasoning version 11 on its Intelligence Index, 14th of 67 open-weights models in the 40B to 150B class; NVIDIA's Nemotron 3.5 Super VL, at a similar 120B, scores 26. Mistral's own comparisons are against gpt-oss-120b, which it says Small 4 matches or beats on AA LCR, LiveCodeBench and AIME25, alongside a self-reported 71.2 on GPQA Diamond.
Where it wins is cost per answer. Artificial Analysis measured 182 output tokens a second on Mistral's API with 0.91 seconds to first token, and ranks it first of eight in its class on cost per task, at about $0.02. Mistral says it reaches 0.72 on the AA long-context test while writing about 1,600 characters, where Qwen models need three to four times as much text, which matters when output tokens cost four times as much as input.
The name undersells the hardware. "Small" here means 119B total parameters: the FP8 checkpoint's own vLLM example splits it across two GPUs and Mistral's stated minimum is four H100s, so this is a server model, not a laptop one like gpt-oss-20b or the smaller Gemma 4 sizes. Apache 2.0 does mean no attribution clause, revenue cap or usage notice, which is more than MiniMax-M3 offers.
On Mistral's own API, free-tier inputs and outputs may be used for training unless the user opts out, and paid users can opt out at any time; zero data retention is a separate request. For a European buyer who wants a French vendor's API now and the option to run identical weights in-house later, that pairing is rarer than the benchmark table suggests.
Where it earns its place
A support-operations lead at a 200-agent European contact centre
Draft first replies and triage tickets in five languages
With reasoning_effort set to none, a million replies a month at 1,500 input and 250 output tokens is about $375 on the API, or about $190 through the batch API for overnight triage. Mistral is an EU vendor, and paid accounts can opt out of training.
A platform engineer at a bank that cannot send documents to a third-party API
Run a vision-capable assistant on in-house GPUs
Serve the FP8 checkpoint with vLLM across two H100-class GPUs, as the model card's example does, or four for Mistral's recommended headroom. Apache 2.0 needs no notice to Mistral and has no revenue threshold, and the same model is on the API for overflow.
A developer adding document Q&A to a budget SaaS product
Answer questions about screenshots and scanned forms
Small 4 reads images directly and supports structured outputs and function calling, so one call can return typed JSON from a screenshot at Small-model prices. Harder reasoning or multi-step coding is better routed to Mistral Medium 3.5 or a larger model, given its index score.
Before you commit
- Setup
- API: call mistral-small-latest or mistral-small-2603 and set reasoning_effort per request. Self-hosted: vllm serve with Mistral's tool-call and reasoning parsers; Transformers support needed its main branch at release.
- Integrations
- Mistral API (chat completions, conversations, function calling, structured outputs, batch), Hugging Face, NVIDIA NIM and build.nvidia.com, OpenRouter; vLLM, SGLang, llama.cpp and Transformers, with fine-tuning through NVIDIA NeMo.
- Data and privacy
- Apache 2.0 allows commercial use with no attribution or revenue limits. On Mistral's API, free-tier inputs and outputs may be used for training unless you opt out; pay-as-you-go users can opt out at any time; thumbs-up and thumbs-down feedback is usable on every plan.
- Cost at scale
- 1M chat replies a month at 1,500 input and 250 output tokens cost about $375 on the API and about $190 in batch. Self-hosting needs two to four H100-class GPUs per replica at FP8, or one DGX B200.
- Limits
- 256K-token context; maximum output is not stated in the docs we read, and no rate limits appear on the model page. High reasoning effort writes far more tokens, which is where cost and latency rise.
What's good
- Apache 2.0 with no attribution or revenue terms
- $0.15/$0.60 API, half that in batch
- 182 tokens/s on Mistral's API (Artificial Analysis)
- Image input and switchable reasoning in one model
What's not
- Scores 11 on Artificial Analysis's index against 26 for Nemotron 3.5 Super VL
- Needs two to four data-centre GPUs despite the name
- Mistral benchmarks it only against gpt-oss-120b and Qwen
- Free API tier may train on data unless you opt out
Best for
Teams wanting a low-cost, Apache-licensed model for chat, documents and images, on Mistral's API or their own GPUs.
Alternatives
Mistral Small 4: common questions
- Is Mistral Small 4 worth it?
- Cheap, fast and cleanly licensed, but independent tests put it well behind the open models it shares a server with. We rate it 3.7 out of 5.
- How much does Mistral Small 4 cost?
- Free weights under Apache 2.0. API: $0.15 input / $0.60 output per 1M tokens; batch 50% off; cached input up to 90% off.
- What are the downsides of Mistral Small 4?
- Scores 11 on Artificial Analysis's index against 26 for Nemotron 3.5 Super VL; needs two to four data-centre GPUs despite the name; mistral benchmarks it only against gpt-oss-120b and Qwen; free API tier may train on data unless you opt out.
- Who is Mistral Small 4 best for?
- Teams wanting a low-cost, Apache-licensed model for chat, documents and images, on Mistral's API or their own GPUs.
- How much does Mistral Small 4 cost for a team or in production?
- 1M chat replies a month at 1,500 input and 250 output tokens cost about $375 on the API and about $190 in batch. Self-hosting needs two to four H100-class GPUs per replica at FP8, or one DGX B200.
- Is Mistral Small 4 safe for company data?
- Apache 2.0 allows commercial use with no attribution or revenue limits. On Mistral's API, free-tier inputs and outputs may be used for training unless you opt out; pay-as-you-go users can opt out at any time; thumbs-up and thumbs-down feedback is usable on every plan.
- What are the best Mistral Small 4 alternatives?
- The closest alternatives we have reviewed are gpt-oss-120b, Nemotron 3.5 Super VL, Mistral Medium 3.5 and Gemma 4.
Answers checked Oct 22, 2026.