---
license: apache-2.0
tags:
- llama.cpp
- chat-template
- qwen3-coder-next
- jinja
- tool-use
---

# Qwen3-Coder-Next llama.cpp chat template

This repository contains a small llama.cpp-oriented patch of the official
`Qwen/Qwen3-Coder-Next` chat template.

It stays close to the upstream Qwen template and only adds compatibility fixes
that are useful for OpenAI-compatible coding agents such as Codex, Aider,
OpenCode-style harnesses, and other clients that may send modern message
shapes.

## Files

- `Qwen3-Coder-Next.jinja` - patched llama.cpp Jinja chat template.
- `smoke_qwen3_coder_next_template.py` - small Jinja2 smoke test.

## Changes from the official template

- Accepts `developer` messages and renders them as Qwen `system` messages.
- Adds `render_content(content)` so string content, typed content arrays,
  `None`, and simple multimodal placeholders do not crash the template.
- Handles tool responses whose `content` is a typed content array.
- Handles `tool_call.arguments` as either a mapping or a pre-serialized string.

The goal is not to redesign Qwen's tool-calling format. The tool prompt and XML
tool-call format are intentionally kept close to Qwen's official template.

## llama.cpp usage

Example `models.ini` section:

```ini
[Qwen3-Coder-Next]
model = /path/to/Qwen3-Coder-Next-UD-IQ3_XXS.gguf
jinja = on
chat-template-file = /path/to/Qwen3-Coder-Next.jinja
chat-template-kwargs = {"enable_thinking": false}
```

## Smoke test

```bash
python -m pip install jinja2
python smoke_qwen3_coder_next_template.py
```

Expected result:

```text
ALL_OK
```

## Notes

This template is intended for llama.cpp and minijinja-style rendering. The smoke
test uses Python Jinja2 only to catch obvious template failures and common
agent-message shapes; final behavior should still be validated with the actual
serving runtime.

Based on the official Apache-2.0 Qwen3-Coder-Next chat template.