<MCP.Playground/>
online

Mistral Small 2603

Runs MCP tools

Mistral’s March 2026 small model, which folds reasoning, vision and agentic coding into one set of Apache 2.0 weights. Reasoning is a per-request switch, so one model can serve both fast and careful tool calls.

Vendor
Mistral AI
Released
17 March 2026
Context
262K tokens
Input
Text, image

Paste a server URL, pick a model, and watch it call your tools in a real conversation. No install.

No server of your own? Leave it blank and use one of the public mock servers.

MCP Evals

Can Mistral Small 2603 actually complete tasks with your server?

One chat shows the tool calls work. An eval writes real tasks from your tool schemas, has Mistral Small 2603 drive each one, and reports where it got the wrong answer even though every call succeeded. Pick Mistral Small 2603 as the driver.

Run evals →See a sample

What Mistral Small 2603 is

Mistral Small 4 is a 119B-parameter mixture-of-experts model with 6.5B active per token: 4 of 128 experts. It replaces three separate lines: Magistral for reasoning, Pixtral for vision and Devstral for agentic coding. It has a 256K window, takes text and image input, and is released under Apache 2.0. Mistral says it cuts end-to-end completion time by 40% and serves three times as many requests per second as Small 3.

What Mistral AI says

  • Matches or beats GPT-OSS 120B on Mistral’s benchmark set while producing shorter output.
  • AA LCR: 0.72 with about 1.6K characters of output, where Mistral says Qwen models need 3.5–4× more output for a similar score.
  • 40% lower end-to-end completion time and 3× the requests per second of Mistral Small 3.

These are the vendor’s own claims, not measurements of ours. Run the model against your own server to find out whether they hold for your tools.

How it reaches MCP

  • Le Chat and Mistral’s Agents API, both of which connect to MCP servers
  • The open weights, served with vLLM, SGLang or llama.cpp, behind any MCP client, including on a single workstation
  • MCP Playground’s Agent Studio: paste your server URL above and run Mistral Small 4 against it in the browser, with no Mistral key

What to watch for

  • reasoning_effort takes none or high. With none it answers straight away, which suits single lookups. With high it plans first, which suits chains where one result feeds the next. Pick per server rather than leaving one setting for everything.
  • Only 6.5B parameters are active per token. It is quick and cheap, but on servers with many similar tools, check that it picks the right one before trusting it.
  • OpenRouter lists tools and tool_choice but not parallel_tool_calls, so expect one call per turn.

Run the complex-schema mock twice on Mistral Small 4, once with reasoning off and once with it on, using the same prompt. configure_workflow takes the most structured input, so it shows most clearly whether the extra thinking gets the arguments right first time. All the mock servers →

Available here

ModelModel ID
Mistral Small 2603mistralai/mistral-small-2603

Go deeper

Frequently asked questions

Is Mistral Small 4 open source?
Yes. The weights are on Hugging Face under Apache 2.0, which allows commercial use.
What is mistral-small-2603?
The API name for Mistral Small 4. 2603 is the release: March 2026.
Does Mistral Small 4 support tool calling?
Yes. It has native function calling and JSON output, and OpenRouter lists tools and tool_choice. MCP servers work through any client that turns MCP tools into function calls.
Can I run Mistral Small 4 locally?
With 6.5B active of 119B total, quantized builds run on a single high-memory GPU or a large Apple silicon Mac, and llama.cpp and LM Studio support it. It is one of the few models here that is realistic to run next to a local MCP server.

Sources

Mistral Small 4 for MCP: Specs & Tool Calling — Test Free | MCP Playground