<MCP.Playground/>
online

MiniMax M3

Runs MCP tools

MiniMax’s open-weight model from June 2026, combining frontier-level coding, a 1M window and native multimodality. MiniMax reports a score on MCP Atlas, a benchmark built on real MCP tool use, so its claims bear directly on your server.

Vendor
MiniMax
Released
31 May 2026
Context
1M tokens
Input
Text, image, video

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.

In your browser

MCP Agent Studio

Server has a public URL? Chat with MiniMax M3 on it, watch every tool call, and compare models side by side.

Open MCP Agent Studio →

On your computer

MCP Playground Desktop

Server runs on localhost or over stdio? The browser can’t reach it. The desktop app can, with MiniMax M3 in Agent Studio. Download free

Mac (Apple Silicon)Mac (Intel)Windowsor Microsoft Store

MCP Evals

Can MiniMax M3 actually complete tasks with your server?

One chat shows the tool calls work. An eval writes real tasks from your tool schemas, has MiniMax M3 drive each one, and reports where it got the wrong answer even though every call succeeded. Pick MiniMax M3 as the driver.

Run evals →See a sample

What MiniMax M3 is

MiniMax M3 succeeds M2.7. MiniMax announced it on 1 June 2026. It is built on MiniMax Sparse Attention (MSA), a new attention design that MiniMax says cuts per-token compute to a twentieth of the previous generation at long context, which is how it reaches a 1M-token window. It was trained on mixed modalities from the start, takes image and video input, and can operate a desktop. Thinking can be switched on or off.

What MiniMax says

  • MCP Atlas: 74.2%.
  • SWE-Bench Pro: 59.0%.
  • Terminal-Bench 2.1: 66.0%.
  • Per-token compute about 1/20 of the previous generation at long context.

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

  • MCP Playground’s Agent Studio: paste your server URL above and run MiniMax M3 against it in the browser, with no MiniMax key
  • MiniMax’s Anthropic-compatible Messages endpoint, so Claude Code and other Anthropic-based MCP clients can point at it with a base-URL change
  • MiniMax’s OpenAI-compatible chat completions endpoint, for MCP clients that accept a custom OpenAI-compatible endpoint

What to watch for

  • It uses interleaved thinking: it reasons between tool calls, not only before the first. MiniMax’s docs say the complete assistant message, thinking blocks included, must go back into history in multi-turn tool loops. A client that strips thinking can break the reasoning chain.
  • MCP Atlas measures real MCP tool use, so the 74.2% is closer to your workload than most benchmarks. It is still MiniMax’s own number on MiniMax’s harness.
  • Its Terminal-Bench 2.1 score (66.0%) is well below the ~86–90% range of DeepSeek V4.1 Flash, Qwen 3.8 Max and Kimi K3. Long terminal-style chains are the place to check it.

Run M3 and M2.7 on the same four-step task against the complex-schema mock: create_user_profile, process_order, analyze_data, then configure_workflow. Interleaved thinking should show as M3 adjusting after each result rather than following a plan fixed at the start. All the mock servers →

Available here

ModelModel ID
MiniMax M3minimax/minimax-m3

Go deeper

Frequently asked questions

Does MiniMax M3 support MCP?
Yes. MiniMax reports 74.2% on MCP Atlas, a benchmark of real MCP tool use. Its API is both Anthropic- and OpenAI-compatible, so clients like Claude Code can drive it. You can run it against your own server in MCP Playground with no MiniMax key.
What is interleaved thinking?
The model reasons again after each tool result instead of planning once and executing a fixed sequence. It helps multi-step tool use, but only if the client sends the thinking blocks back with the conversation.
Is MiniMax M3 open weight?
Yes. At launch MiniMax said the weights and technical report would follow within ten days. MiniMax calls it the first open-weight model to combine frontier coding, 1M context and native multimodality.
What changed from MiniMax M2.7?
M3 moves to the new MSA attention, raises the window from about 205K to 1M tokens, and adds image and video input. M2.7 was text-only. Both are available here if you want to compare them on your tools.

Sources