What is Mistral Small 4?
Mistral Small 4 is a 119B-parameter mixture-of-experts model with 128 experts, 4 of them active per token, so about 6.5B parameters do the work on each step. It folds Mistral's Magistral reasoning, Pixtral vision and Devstral coding models into one set of weights. It reads text and images, writes text and handles 256K tokens, and a reasoning_effort setting switches between quick replies and step-by-step reasoning on each request.
It is Mistral's current Small model: Mistral points users of the retired Small 3.2 and Magistral Small to it, while the pricier Mistral Medium 3.5 sits above it.
What changed from Mistral Small 3.2?
It grew from a 24B dense model into a 119B mixture of experts that also covers reasoning and coding, with twice the context.
- Size
- Small 3.2 was a 24B dense model; Small 4 has 119B parameters with about 6.5B active per token.
- Reasoning
- Built in and switchable per request, replacing the separate Magistral Small models.
- Context
- 256K tokens, up from 128K on Small 3.2.
- Speed
- Mistral measures 40% lower completion time and three times more requests per second than Mistral Small 3.
Who is Mistral Small 4 best for?
Mistral Small 4 is best for teams that want one open model for chat, coding agents, reasoning and image input at a low API price, with reasoning they can turn up or down on each request.
One model instead of three
It covers chat, reasoning, coding agents and image questions, so a single deployment replaces separate instruct, Magistral and Devstral models.
Apps that need speed and depth
A reasoning_effort setting switches between quick replies and step-by-step reasoning on each request.
Long documents and images
A 256K-token context and image input suit document parsing, extraction and visual analysis.
Self-hosting on data-centre GPUs
The Apache 2.0 weights run on vLLM, llama.cpp, SGLang and Transformers, and Mistral lists four H100 GPUs as the minimum setup.
What's included with Mistral Small 4?
Per-request reasoning, image input, tool calls and JSON output, speculative decoding and NVFP4 builds for self-hosting, half-price batch jobs, and free trials on NVIDIA's site.
- Reasoning on demand
- Set reasoning_effort to none for fast replies or to high for a full thinking step before the answer, per request.
- Image input
- Reads images alongside text for document parsing and visual questions; output is text only.
- Tool calls and structured output
- Native function calling and JSON output, plus Mistral's Agents API, built-in tools and predicted outputs.
- Faster self-hosting
- Mistral publishes a small EAGLE speculative decoding model and an NVFP4 checkpoint to speed up inference on your own GPUs.
- Batch and Priority tiers
- Batch jobs cost half, at $0.075 / $0.3 per million tokens; the Priority tier costs $0.263 / $1.05 for committed service levels.
- Free trials on NVIDIA
- Developers can try it free on build.nvidia.com, and it ships as an NVIDIA NIM container for production.
How much does Mistral Small 4 cost?
Mistral Small 4 is free to download and run under Apache 2.0, and on Mistral's API it costs $0.15 per million input tokens and $0.6 per million output tokens.
- Batch: $0.075 / $0.3
- Priority: $0.263 / $1.05
- Cached input tokens cost 90% less
These are Mistral's own API prices in US dollars. EU or global regional endpoints add 10% on supported models. The weights are free on Hugging Face under the Apache 2.0 license, Mistral's free plan includes $10 a month in API credits, and NVIDIA's build.nvidia.com lets developers try the model for free. Full pricing on mistral.ai. We checked these prices on .
How do readers rate Mistral Small 4?
No one has rated Mistral Small 4 yet. Be the first.
Sign in to rate. One rating per account.
Mistral Small 4 pros and cons
It is cheap, quick and sparing with tokens, with Apache 2.0 weights, but at 119B it is too big for most home GPUs and it trails Qwen3.5-122B on many benchmarks.
Pros
- Low price for its class: $0.15 / $0.6 per million tokens, which several Hacker News users call a steal and among the best value in its bracket.
- Short answers: Mistral's charts show it matching Qwen's scores on AA LCR with 3.5 to 4 times less output, and users say Qwen spends far more tokens reasoning on the same work.
- Fast for its size, since only 6.5B parameters are active per token, and it comes with a speculative decoding model that users expect to speed it up further.
- Apache 2.0 weights from a European lab, which users say matters to companies that avoid US-hosted or Chinese models.
Cons
- Not small: at 119B parameters, users report 4-bit files of about 72GB, and Mistral's minimum for serving it is four H100 GPUs, so single consumer GPUs cannot run it.
- It trails Qwen3.5-122B: users reading Mistral's own charts note it does not beat Qwen overall, and some found it behind Qwen at similar sizes in their own tests.
- Mixed on creative writing: some users say it invents details and ignores formatting rules in long role-play scenes, though others rate it on par with DeepSeek there.
Is Mistral Small 4 open source?
Yes. Mistral Small 4's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Mistral Small 4 alternatives?
Mistral Small 4 is an open source alternative to Nemotron 3 Super. NVIDIA Nemotron 3 Ultra and Qwen3.6-27B do the same job and are open source too.
Nemotron 3 SuperFreeChoose it when you need up to 1M tokens of context in a model of the same 120B size and text-only input is fine; NVIDIA lists two H100 80GB GPUs as the minimum for its FP8 version.
NVIDIA Nemotron 3 UltraOpen SourceNVIDIA's 550B open-weight reasoning model with a 1M context.
Qwen3.6-27BOpen SourceAlibaba's 27B open coding and vision model.
Gemma 4Open SourceChoose it when you must run on one consumer GPU: Google puts its 31B model at about 17.5GB in 4-bit, and Gemma 4 is free on the Gemini API's free tier.
DeepSeek V4.1 FlashOpen SourceChoose it when you want a 1M-token context and up to 384K tokens of output through an API; it costs the same $0.15 / $0.60 off-peak and $0.30 / $1.20 at peak.
Questions about Mistral Small 4
What is Mistral Small 4?
Mistral Small 4 is Mistral AI's open-weight model that merges its instruct, reasoning and coding models into one, with image input, a 256K context and Apache 2.0 weights, for teams wanting one cheap model.
Who is Mistral Small 4 for?
Mistral Small 4 is best for teams that want one open model for chat, coding agents, reasoning and image input at a low API price, with reasoning they can turn up or down on each request.
Is Mistral Small 4 free?
Yes, Mistral Small 4's weights are free to download and use under Apache 2.0. Calling it on Mistral's API costs $0.15 per million input tokens and $0.6 per million output tokens, and batch jobs cost half that.
How much does Mistral Small 4 cost?
Mistral Small 4's paid plans are Mistral Small 4 API at $0.15 / $0.6 per million input / output tokens.
Is Mistral Small 4 open source?
Yes. Mistral Small 4's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Mistral Small 4 alternatives?
The closest alternatives on Recaply are Nemotron 3 Super, NVIDIA Nemotron 3 Ultra, Qwen3.6-27B, Gemma 4 and DeepSeek V4.1 Flash.
Where can I try Mistral Small 4?
On its official website, mistral.ai.
What hardware does Mistral Small 4 need?
Mistral lists a minimum of 4x NVIDIA HGX H100, 2x HGX H200 or 1x DGX B200 to self-host Mistral Small 4. It has 119B parameters in total but only about 6.5B active per token, spread over 128 experts.
Why is Mistral Small 4 called small if it has 119B parameters?
Mistral names it for its place in the lineup, not its file size. Only about 6.5B parameters work on each token, so it answers quickly and costs little per token, but all 119B still have to fit in memory. It replaces the 24B Mistral Small 3.2.
How do I turn on reasoning in Mistral Small 4?
Set reasoning_effort to "high" in the request for step-by-step reasoning, or to "none" for fast replies. With high, the response includes a thinking chunk before the answer and uses more tokens; Mistral suggests a temperature of 0.7 in that mode.
Mistral Small 4 vs Mistral Medium 3.5: which should I use?
Small 4 is much cheaper at $0.15 / $0.6 per million tokens against $1.5 / $7.5, and its licence is Apache 2.0. Medium 3.5 is a 128B dense model Mistral built for long coding and agent jobs, scoring 77.6% on SWE-Bench Verified, under a modified MIT licence.
Our verdict
Mistral Small 4 is a good default open model for teams that host on data-centre GPUs or want a cheap API with reasoning on tap.
It covers chat, coding, reasoning and images in one Apache 2.0 model at $0.15 / $0.6 per million tokens and writes shorter answers than Qwen, though Qwen3.5-122B still scores higher on many benchmarks.
- Skip it for
- Setups with one consumer GPU, where Gemma 4 or Ministral 3 fit better.
- Check first
- Run your own prompts with reasoning_effort set to both none and high, since quality and token use change a lot between them.
Mistral Medium 3.5
Mistral 3
Qwen3.5 Small
Step 3.7 Flash
DeepSeek-V3.2-Speciale
Ring-2.6-1T