Tool use does not work with OpenAI compatible endpoint running Qwen3-coder

[regunakyle]

Hi, I have a self hosted OpenAI compatible server (tabbyAPI) that runs Qwen3-coder. I set the provider to openai in config.

Unfortunately tool call does not work with this. I suspect the cause is that the tool call template of OpenAI differs from Qwen's.

Is there a way to tell Continue that my endpoint is running a specific supported model? (as written in the docs)

Tool call works with the real OpenAI endpoint and Anthropic.

[bdougie]

The Issue

The issue you're experiencing with Qwen3-Coder:30B Instruct and tool calling through a self-hosted OpenAI-compatible endpoint is related to how Continue determines which tool-calling format to use.

How Continue Handles Tool Calling

According to the Agent Mode Model Setup documentation, Continue uses two approaches for tool calling:

1. Native tool-calling APIs - Uses the provider's built-in tool support
2. System message tools - Embeds tools as XML in the system message

Continue automatically detects which approach to use based on the model's capabilities. However, when you configure provider: openai for your tabbyAPI endpoint, Continue assumes it's communicating with an actual OpenAI model and expects OpenAI's tool-calling format.

The Problem

Qwen3-Coder:30B Instruct uses a different tool-calling template than OpenAI. When Continue sends requests in OpenAI's format to your endpoint, Qwen3-Coder:30B Instruct doesn't understand the structure, causing tool calls to fail.

Recommended Solution

You should configure your model with provider: openai but remove the tool_use capability to force Continue to use system message tools instead of native OpenAI tool APIs:

models:
  - name: Qwen3-Coder
    provider: openai
    model: Qwen3-Coder:30B Instruct
    apiBase: http://your-tabbyapi-endpoint
    apiKey: your-key
    capabilities: []  # Explicitly remove tool_use capability

Or alternatively, don't specify any capabilities at all and Continue should fall back to system message tools:

models:
  - name: Qwen3-Coder
    provider: openai
    model: Qwen3-Coder:30B Instruct
    apiBase: http://your-tabbyapi-endpoint
    apiKey: your-key
    # No capabilities specified - Continue will use system message tools

This configuration tells Continue to use the XML-based system message tools approach, which should work with Qwen3-Coder:30B Instruct since it can follow instructions to generate structured XML output.

Alternative Approach

If the above doesn't work, you might also try setting the provider to ollama or another provider type that defaults to system message tools, though this may require adjusting your endpoint configuration.

Note on Configuration

The valid model capabilities in Continue are:

• tool_use - Enables function/tool calling support
• image_input - Enables image upload and processing support

For more details, see the Model Capabilities documentation.

Would you be able to test this configuration and report back whether it resolves the tool-calling issues?

[regunakyle]

Thanks, it worked! Closing this issue now.