Qwen 3.8 Max

Runs MCP tools

Alibaba’s largest Qwen model, released 3 August 2026: 2.4 trillion parameters, multimodal from the start, with open weights. Alibaba aims it at autonomous software engineering and long-horizon work.

Vendor
Alibaba (Qwen)
Released
3 August 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.

MCP Evals

Can Qwen 3.8 Max actually complete tasks with your server?

One chat shows the tool calls work. An eval writes real tasks from your tool schemas, has Qwen 3.8 Max drive each one, and reports where it got the wrong answer even though every call succeeded. Pick Qwen 3.8 Max as the driver.

What Qwen 3.8 Max is

Qwen 3.8 Max is a 2.4T-parameter mixture-of-experts model. Its open-weight release, Qwen3.8 2.4T A95B, activates 95B parameters per token. It takes text, image and video input, has a 1M-token window with up to 131K tokens of output, and a thinking budget of up to 262K tokens. Alibaba followed the launch with a 0902 snapshot on 2 September, and that updated snapshot is the one that runs here.

What Alibaba (Qwen) says

  • Terminal-Bench 2.1: 86.6, against 84.6 for Claude Opus 4.8 and Claude Fable 5 and 88.8 for GPT-5.6 Sol.
  • OSWorld-Verified: 86.1.
  • SWE-bench Pro: 67.7. FrontierSWE: 73.5.
  • GPQA Diamond: 92.6.

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 Qwen 3.8 Max against it in the browser, with no Alibaba Cloud key
  • Qwen Code, Alibaba’s open-source terminal agent, which acts as an MCP client
  • Alibaba Cloud Model Studio’s OpenAI-compatible API, so MCP clients that accept a custom OpenAI-compatible endpoint can point at it with a base-URL change

What to watch for

  • The model behind the name has moved. The 0902 snapshot replaced the August launch model, so results you recorded in August may not match a run today. Re-test before comparing.
  • Its 86.1 on OSWorld-Verified is a computer-use score. MCP servers that return screenshots or UI state as image content are where that strength should show, if your client passes images through.
  • Qwen’s own API ships built-in tools such as code_interpreter and web_search. Those are separate from your MCP tools. If both are enabled, check which one the model chose before you blame your server.

Run the complex-schema mock on Qwen 3.8 Max and Qwen 3.7 Max with the same prompt. configure_workflow takes the most structured input of the four tools, which makes it the clearest place to see whether the larger model gets the arguments right first time. All the mock servers →

Available here

ModelModel ID
Qwen 3.8 Maxqwen/qwen3.8-max

Go deeper

Frequently asked questions

Is Qwen 3.8 Max open source?
Alibaba released open weights shortly after the 3 August launch, as Qwen3.8 2.4T A95B. At 2.4T parameters it needs a multi-node datacenter setup to run. To try it against an MCP server without that, use it here.
What is Qwen 3.8 Max 0902?
An updated snapshot of Qwen 3.8 Max, released 2 September 2026. It is the version that runs when you pick Qwen 3.8 Max in MCP Playground.
Does Qwen 3.8 Max support MCP?
Yes. Qwen Code is an MCP client, Alibaba’s API is OpenAI-compatible so MCP servers work through any client that translates tools into function calls, and you can run it against your own server in MCP Playground.
Qwen 3.8 Max or Kimi K3?
Both are open-weight trillion-parameter models from China with 1M windows. Alibaba reports 86.6 on Terminal-Bench 2.1 and Moonshot 88.3 for K3. Run the same prompt on both against your server here and compare the tool calls.

Sources

Qwen 3.8 Max for MCP: Specs & Tool Calling — Test Free | MCP Playground