AI model by Mistral AI
Mistral Small 4 for business
Mistral's efficient open-weight model under Apache 2.0, one hybrid model for instructions, reasoning and coding, with tool calling and a 256K window.
- Open weights
- General purpose
- Coding
- Open weights
- Reviewed on September 24, 2026
- Mistral AI
Capability tiersRelative, not benchmarks
Key facts about Mistral Small 4
- Model ID at review
- mistral-small-2603
- Provider
- Mistral AI
- Main category
- Open weights
- Open weights
- Yes, Apache 2.0
- Last reviewed
- September 24, 2026
Fit
Where it fits and where it does not
Good at
-
Efficient serving, with about 6.5 billion active parameters.
-
Instructions, reasoning and code in one model.
-
Tool calling and structured outputs.
-
Self-hosting under Apache 2.0.
-
Long inputs, with a 256K token window.
Not the right choice for
-
The hardest reasoning tasks.
-
Image input, which we could not confirm at review time.
-
Teams that do not want to run infrastructure, unless they use the API.
-
Use without testing on your own data.
Use cases
Business use cases we would use it for
Private processing
Code helpers
Our notes
When we would choose it
Mistral Small 4 was released in March 2026 under the Apache 2.0 license. Mistral describes it as one hybrid model for instruction following, reasoning and coding. It is a mixture-of-experts model with 119 billion parameters in total and about 6.5 billion active, which makes it efficient to serve for its capability. It supports tool calling, structured outputs and a 256K token context window.
For self-hosted projects, Small 4 is one of the most practical options at review time. The permissive license, efficient serving and support for tools and JSON output cover what most business assistants and automations need.
We could not confirm image input from the model card at review time, so the vision meter is set low. If you need to read images or scans, check the current documentation or look at Gemma 4 or Mistral Large 3.
The model is also available through the Mistral API, which is a simple way to test it before deciding whether to host it. We would run your real tasks through it and a hosted leader, then weigh the quality difference against the benefits of running it yourself. See open weights models.
Keep in mind that total size, not active size, decides how much memory you need. A mixture-of-experts model is fast per token but still has to be loaded in full. Size your servers on the total, and test throughput with realistic traffic before launch.
Before you commit
Things to check before you commit
-
Memory
Total size is 119 billion parameters, so memory needs are higher than the active count suggests.
-
Vision
Confirm image support in the current model card if you need it.
-
Quality
Compare with Gemma 4 and hosted models on the same test set.
-
License
Apache 2.0 is permissive. Confirm it fits your use.
Alternatives
Models to compare it with
Gemma 4
Google's open-weight model family under Apache 2.0, in sizes from phone-friendly to 31B, with image input and function calling.
Mistral Large 3
Mistral's open-weight general-purpose flagship under Apache 2.0, with image input, tool calling, structured outputs and a 256K token window.
Qwen3-Coder
Qwen's open-weight coding models under Apache 2.0, from the efficient Qwen3-Coder-Next to the large 480B model, with tool use and long context.
Keep exploring
Solutions, services and guides
Related services
View all related services- LLM integration Add a large language model to software you already have, with the guardrails, costs and logging handled.
- AI model evaluation Test candidate AI models on your real data and tasks, then pick the one that balances quality, speed and cost.
- AI Product Development AI agents, knowledge assistants, copilots and document automation built into the way your team already works.
- AI agents AI that takes actions in your systems, such as qualifying leads or processing requests, with people checking the results.
- AI copilots AI helpers built into your product or internal tools that draft, summarize and suggest while people decide.
- AI workflows Step-by-step automations where AI handles the reading, sorting and drafting inside a process you control.
Solutions
View all solutions- Internal help desk assistant An assistant in Slack or Teams that answers policy and how-to questions from your handbooks and opens tickets when needed.
- Email triage Sort a shared inbox by topic and urgency, pull out the key details and draft replies for a person to send.
- AI-assisted content workflows Briefs, first drafts, edits and repurposing in a workflow where people set the angle and approve every word.
- AI translation workflows Translate websites, products and support content quickly with AI, a shared glossary and native-speaker review where it matters.
- Feedback analysis Read every review, survey and ticket, group them by theme and sentiment, and show what customers keep asking for.
- Screening assistant Summarize each application against your criteria, flag strong matches and missing information, and leave every decision to a person.
Industries
View all industriesCase studies
View all case studiesGuides and articles
View all guides and articles- How to choose an AI model for your business A step-by-step way to pick an AI model by testing candidates on your own tasks, data rules and budget.
- RAG vs fine-tuning Two ways to make AI work with your knowledge, compared by cost, accuracy and upkeep.
- What is RAG and when do you need it? How retrieval-augmented generation lets AI answer from your documents, when it fits and how to build it well.
- How to keep customer data safe when using AI Practical steps to protect customer data when you use AI services, from data terms to access and logging.
- Prompt writing basics for business teams How to write prompts that get consistent, useful results from AI tools, with templates your team can reuse.
MCP servers
View all mcp serversGlossary terms
View all glossary terms- Large language model A large language model, or LLM, is an AI model trained on vast amounts of text that can understand and generate language, and often images and code.
- Machine learning Machine learning is a way of building software that learns patterns from data to make predictions or decisions, instead of following only hand-written rules.
- Fine-tuning Fine-tuning is further training an existing AI model on your own examples so it learns a specific style, format or task.
- Inference Inference is the step where a trained AI model is used to produce an output, such as an answer, a label or a prediction, from new input.
- Token A token is a small piece of text, often part of a word, that AI models read and write, and that providers use to measure limits and pricing.
- Zero-shot classification Zero-shot classification is sorting text or images into categories an AI model was not specifically trained on, using only the category names or descriptions.
FAQ
Questions people ask us
Have a question that is not here? Ask us directly.
It is released under Apache 2.0, a permissive license. Read the license for your case.
It has about 6.5 billion active parameters but 119 billion in total, so plan memory for the full model.
Yes. It is available through the Mistral API.