Qwen 3.8 27B for MCP: Specs & Tool Calling — Test Free

What Qwen 3.8 27B is

Qwen 3.8 27B is a dense 27B-parameter vision-language model. It mixes Gated DeltaNet linear-attention layers with Gated Attention layers, three to one, which keeps a long window affordable: 262K natively and up to 1M with YaRN scaling. It takes text, image and video input. Thinking is on by default at xhigh effort, and preserved thinking is on too, so earlier reasoning stays in the history across turns.

What Alibaba (Qwen) says

Terminal-Bench 2.1: 73.0.
SWE-bench Pro: 61.7.
OSWorld: 84.3.
GPQA Diamond: 89.2. LiveCodeBench: 90.3.

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

The open weights in LM Studio, Ollama, llama.cpp or vLLM, behind any local MCP client. The whole setup can run on one machine.
Qwen Code, Alibaba’s open-source terminal agent, which acts as an MCP client
MCP Playground’s Agent Studio: paste your server URL above and run Qwen 3.8 27B against it in the browser, with nothing to download

What to watch for

The default reasoning effort is xhigh, and it shows. Simon Willison measured 22,276 reasoning tokens for 3,223 tokens of output on one prompt, and recommends low or no reasoning. Each MCP tool call waits for that thinking, so set reasoning_effort to low or medium before judging it on your server.
Run locally, it generated 15–30 tokens a second in Willison’s tests. That is fine for checking tool calls are right, and slow for long chains.
OpenRouter lists tools, tool_choice and structured_outputs, but not parallel_tool_calls. Expect one call per turn.

Run the same four-step task against the complex-schema mock on Qwen 3.8 27B and Qwen 3.8 Max. The 27B is a fraction of the size. If it makes the same calls, it is the one to run locally beside your server while you develop. All the mock servers →

Available here

Model	Model ID
Qwen 3.8 27B	`qwen/qwen3.8-27b`

Go deeper

Qwen 3.8 Max · Alibaba (Qwen)
Gemma 4 31B · Google DeepMind
Mistral Small 2603 · Mistral AI
Every model family that works with MCP

Frequently asked questions

Can Qwen 3.8 27B run locally?

Yes. The Q4_K_M build is about 17GB and runs in LM Studio on a high-memory Mac or a single workstation GPU. The weights are Apache 2.0.

How do I stop Qwen 3.8 27B overthinking?

Set reasoning_effort to low or medium, or turn thinking off with enable_thinking set to false. The default is xhigh, which is slow for tool loops.

Does Qwen 3.8 27B support tool calling?

Yes. OpenRouter lists tools and tool_choice, and Simon Willison ran a full coding agent on it with reliable tool calls. MCP servers work through any client that turns MCP tools into function calls.

Qwen 3.8 27B or Qwen 3.8 Max?

Max is the 2.4T flagship; the 27B is the one you can run yourself. Qwen reports 73.0 on Terminal-Bench 2.1 for the 27B and 86.6 for Max. Both are available here.

Qwen 3.8 27B