Prerequisites

• I am running the latest code. Mention the version if possible as well.
• I carefully followed the README.md.
• I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• I reviewed the Discussions, and have a new and useful enhancement to share.

Feature Description

Here is a short explanation of the differences between the chat templates for Qwen3-Coder and Qwen3-Instruct, with clear examples.

The primary difference is how they format tool calls:

• Qwen3-Coder uses a completely custom XML format for both defining and calling tools.
• Qwen3-Instruct uses standard JSON for tool definitions and expects JSON output from the model, wrapped in a simple <tool_call> tag.

1. Qwen3-Coder Template (Custom XML Format)

This template instructs the model to generate tool calls using unique XML tags for the function name and each parameter [2]. This format is not standard and requires a special parser to convert the XML into a usable JSON object [3].

Example Tool Call Output:

<tool_call>
  <function=search_products>
    <parameter=query>
      waterproof running shoes
    </parameter>
    <parameter=sort_by>
      price_low_to_high
    </parameter>
  </function>
</tool_call>

Key Takeaway: The entire call is structured with custom XML tags like <function=...> and <parameter=...> [2]. This is why it fails in systems that expect JSON, causing parsing errors [1].

2. Qwen3-Instruct Template (JSON in XML Format)

This template provides the tool definitions to the model as standard JSON. It then instructs the model to respond with a JSON object containing the function name and arguments, all inside a simple <tool_call> wrapper [4].

Example Tool Call Output:

<tool_call>
{"name": "search_products", "arguments": {"query": "waterproof running shoes", "sort_by": "price_low_to_high"}}
</tool_call>

Motivation

Well, don't we want native tool call support for coder models?

Possible Implementation

https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/blob/main/qwen3coder_tool_parser.py

They have a custom parser that converts the model output to JSON.