Qwen3.5 not in tool calling model list

[tyh-carl]

System Info / 系統信息

Qwen3.5 should be able for tool calling. However, when I call it via xinference, it shows the error below:

Error Code: 400
Message: litellm.BadRequestError: XinferenceException - Error code: 400 - {'detail': "Only ['qwen1.5-chat', 'qwen1.5-moe-chat', 'qwen2-instruct', 'qwen2-moe-instruct', 'qwen2.5-instruct', 'qwen2.5-coder-instruct', 'XiYanSQL-QwenCoder-2504', 'QwQ-32B', 'qwen3', 'HuatuoGPT-o1-Qwen2.5', 'DianJin-R1', 'Qwen3-Thinking', 'Qwen3-Instruct', 'Qwen3-Coder', 'Qwen3-VL-Instruct', 'Qwen3-VL-Thinking', 'Qwen3-Next-Instruct', 'Qwen3-Next-Thinking', 'Qwen3-Omni-Instruct', 'Qwen3-Omni-Thinking', 'MiniMax-M2', 'glm4-chat', 'glm4-chat-1m', 'glm-4.5', 'glm-4.5v', 'GLM-4.6', 'GLM-4.7', 'GLM-4.7-Flash', 'llama-3.1-instruct', 'HuatuoGPT-o1-LLaMA-3.1', 'deepseek-v3', 'deepseek-r1-0528', 'Deepseek-V3.1'] support tool calls"}. Received Model Group=Qwen3.5-27B-4bit Available Model Group Fallbacks=None

Running Xinference with Docker? / 是否使用 Docker 运行 Xinfernece？

• docker / docker
• pip install / 通过 pip install 安装
• installation from source / 从源码安装

Version info / 版本信息

v2.2.0

The command used to start Xinference / 用以启动 xinference 的命令

xinference-supervisor -H MY_IP -p 9997 --supervisor-port 9996 --log-level debug

xinference-work -e "MY_IP" -H "MY_IP" --log-level debug

Reproduction / 复现过程

1. Register mlx-community/qwen3.5-27b-4bit model as qwen3.5 model family
2. launch model
3. it success when using completion mode, but error in agent mode using langgraph.

Expected behavior / 期待表现

qwen3.5 works on langgraph agent mode

[qinxuye]

@llyycchhee 帮忙看下。

[Jin-Cheng-Ming]

xinference v2.3也同样出现这个问题。

[la1ty]

#4703 seems to solve the 400 issue. However, the model returns an xml format of tool_call, which cannot be used directly by coding agents.

[Jin-Cheng-Ming]

#4703 seems to solve the 400 issue. However, the model returns an xml format of tool_call, which cannot be used directly by coding agents.

Yes, I noticed that too.
是呀，我也发现了。

[qinxuye]

So, the tool call might require parsing logic.