The short version. Mistral released Mistral Small 4, a single model that combines reasoning, multimodal (text and image) input, and code generation. It is a Mixture of Experts architecture with 128 experts and 6 billion active parameters per token (119B total). 256k context window. Apache 2.0 licensed.
Why it matters. Before Small 4, Mistral users had to pick between Magistral (reasoning), Devstral (coding), or Mistral Small 3.2 (instruct). Now one model handles all three. Plus 40 percent faster end-to-end completion and 3x more requests per second compared to Small 3.
The interesting thing about Mistral Small 4 is the unification, not the size. Three specialized models collapsed into one. That is the pattern the rest of the open-weight world will probably follow.
Mistral AI shipped Mistral Small 4 this week. The headline is the architecture, not the parameter count: Small 4 is a single model that consolidates three capabilities that previously required three different Mistral models. Reasoning. Multimodal input. Code generation. One model, one API call, one set of weights to host.
This guide breaks down what changed, how it compares to GPT-OSS 120B (the only other notable open-weight model in the same class), and whether non-technical users should care.
What is Mistral Small 4?
It is Mistral’s latest open-weight model, released under Apache 2.0 (the same license as the previous Mistral generations). The architecture is Mixture of Experts: 128 expert sub-networks, with only 4 active per token. Total parameters are 119 billion; active parameters per inference are 6 billion. That structure delivers larger-model quality at smaller-model inference cost.
Three capabilities ship in one model:
- Reasoning. A new
reasoning_effortparameter lets developers toggle between fast lightweight responses and deeper step-by-step reasoning, in the same model. - Multimodal. Native text + image input. Document parsing, screenshot analysis, visual reasoning.
- Code generation. Outperforms GPT-OSS 120B on LiveCodeBench while producing 20 percent fewer output tokens.
How does it compare to other open models?
vs. GPT-OSS 120B (OpenAI’s open-weight model). Small 4 beats it on LiveCodeBench with 20% fewer output tokens. The MoE architecture means it runs faster per request.
vs. Llama 4 (Meta). Llama 4 is dense; Small 4 is sparse MoE. For self-hosters, Small 4 needs less active compute per inference but more total VRAM for weights.
vs. closed models (Claude Opus, GPT-5, Gemini Ultra). Small 4 is positioned in the mid-tier. Not a frontier-quality model. It is a strong open-weight option for teams who need to host their own AI.
For a broader comparison of all the major models, see ChatGPT vs Claude vs Gemini.
Do regular users need to care about this?
Probably not directly. Mistral Small 4 is hosted-or-self-hosted infrastructure, not a consumer chat app. Most readers will encounter Small 4 indirectly: a product they use will quietly run on it. The Mistral Le Chat app (now called Mistral Vibe) will use it; some open-source coding agents will use it; some enterprise deployments will switch to it from older Mistral models.
Two groups should pay attention directly. Developers building AI products who want an Apache-licensed alternative to closed APIs. And teams in regulated industries (healthcare, finance, EU compliance) who must self-host for data sovereignty reasons. Both groups now have a stronger option.
What about hardware to run it?
Modest by frontier-model standards, but still serious hardware. Mistral lists three minimum configurations: 4x NVIDIA H100, 2x H200, or 1x DGX B200. Recommended setups go higher for production throughput.
If you do not own that hardware, every major cloud provider hosts Mistral models. The Mistral API, Mistral AI Studio, Hugging Face Inference, and NVIDIA NIM containers all ship Small 4. Prototyping is free on NVIDIA’s accelerated computing tier; production usage is metered like any other API.
Why unify reasoning, multimodal, and code into one model?
Operational simplicity. The previous Mistral lineup forced teams to manage three different models with three different fine-tuning schedules and three different deployment pipelines. That is overhead. A single model that toggles between modes via a parameter cuts that overhead by two-thirds.
The trade-off is that a generalist model sometimes underperforms specialists on narrow benchmarks. Mistral is betting that the productivity gain from one-model operations outweighs the small per-task loss. Anthropic and OpenAI have made similar bets with Claude 4.x and GPT-5 (single model, multiple modes).
Where does this leave Mistral?
Mistral has been one of the few European AI labs to consistently ship competitive open-weight models. Small 4 strengthens that position. Combined with the Mistral 3 flagship release earlier this year and the Voxtral TTS model, Mistral now has a full stack: large model, mid model, voice model, agent product (see the Mistral Vibe writeup).
For non-technical readers, the takeaway is simpler: there is now a credible third option to ChatGPT and Claude, with European licensing, open weights, and similar mid-tier quality. That competition is good for everyone using AI tools.
Frequently asked questions
What is Mixture of Experts?
An architecture where the model contains many specialized sub-networks (experts), but only a few are active for any given token. Reduces inference cost while keeping total knowledge high. See the AI Glossary for fuller definitions.
What is Apache 2.0 licensing?
A permissive open-source license that allows commercial use, modification, and redistribution. Apache-licensed AI weights can be used by enterprises without negotiating a custom contract.
Can I use Mistral Small 4 in ChatGPT?
No. Mistral models are not available inside ChatGPT. To try Small 4 you can use the Mistral Le Chat / Vibe app, the Mistral API, or any third-party tool that supports Mistral models.
What is the reasoning_effort parameter?
A control developers can set per request. Low values give fast cheap responses; high values trigger step-by-step reasoning. Trade-off between speed and depth, in a single model.
Is Mistral Small 4 better than GPT-4 or Claude?
No. It is mid-tier. The frontier models from OpenAI, Anthropic, and Google are still ahead on quality. Small 4’s value is being open-weight and self-hostable at a competitive mid-tier quality level.
Get the daily Beginners in AI newsletter
One issue a day, plain English. Covers every major model release with the architecture and trade-offs explained for non-engineers.
Get Smarter About AI Every Morning
Free daily newsletter. Built for people who want to use AI well, not chase every model.
Free forever. Unsubscribe anytime.
Sources
- Mistral: Mistral Small 4 announcement
- Mistral on Hugging Face (model weights)
- Mistral API documentation
- Apache 2.0 license text
You might also like
- Mistral AI: complete beginner’s guide
- ChatGPT vs Claude vs Gemini
- AI Glossary
- AI Should Make Us Better
- Best AI Prompts by Job and Use Case
- How to Use Claude AI
Two ways to go further
The AI Prompt Library
1,000+ ready-to-use prompts for Claude, ChatGPT, and Gemini. Stop staring at a blank box.
Get it for $39 →2-Hour Live AI Crash Course
A private, beginner-friendly session across Claude, ChatGPT, Gemini, and the wider landscape.
Book for $125 →