---
title: qwen3-enhanced-chat-template
canonical_url: "https://www.modelscope.cn/models/MaricZg/qwen3-enhanced-chat-template"
md_url: "https://www.modelscope.cn/models/MaricZg/qwen3-enhanced-chat-template.md"
repository: MaricZg/qwen3-enhanced-chat-template
chinese_name: "Qwen3.6 增强版 Chat Template"
last_updated: 2026-05-11
license: mit
library_name:
  - pytorch
frameworks:
  - pytorch
downloads: 10
stars: 1
tags:
  - jinja
  - chat-template
  - qwen
  - qwen3.6
  - lm-studio
  - llama.cpp
  - vllm
  - tool-calling
  - thinking
  - multimodal
---

# qwen3-enhanced-chat-template

> qwen3-enhanced-chat-template - MaricZg 在 ModelScope 开源的模型。Enhanced Chat Template for Qwen3.6

MaricZg/qwen3-enhanced-chat-template 是 ModelScope 魔搭社区上的机器学习模型，采用 mit 许可。

- **Repository**: MaricZg/qwen3-enhanced-chat-template
- **License**: mit
- **Tags**: jinja, chat-template, qwen, qwen3.6, lm-studio, llama.cpp, vllm, tool-calling, thinking, multimodal
- **Downloads**: 10
- **Stars**: 1
- **Last updated**: 2026-05-11

Source: https://www.modelscope.cn/models/MaricZg/qwen3-enhanced-chat-template

---

# Enhanced Chat Template for Qwen3.6

> Based on [froggeric/Qwen-Fixed-Chat-Templates](https://modelscope.cn/models/froggeric/Qwen-Fixed-Chat-Templates) with additional hardening.

Drop-in Jinja2 template that fixes **9 bugs** from the official Qwen3.6 template and adds **dynamic thinking toggle** and **defensive type checking**.

## What's Fixed

| # | Bug | Impact |
|---|-----|--------|
| 1 | No `item is not mapping` guard | C++ engines (llama.cpp/LM Studio) crash |
| 2 | `add_vision_id` undefined access | Strict engines crash |
| 3 | `developer` role rejected | OpenAI-compatible API fails |
| 4 | Missing `video_url` key | Video input ignored |
| 5 | `raise_exception` on tool-only chains | Agent frameworks crash (LangChain/AutoGen/Claude Code) |
| 6 | Solo `<think` tag not handled | Parsing errors |
| 7 | Empty thinking blocks rendered | Wastes context tokens |
| 8 | Tool call missing name check | Malformed output |
| 9 | Tool call arguments not validated | Silent corruption |

## New Features

- **Dynamic thinking toggle**: Insert `<|think_on|>` / `<|think_off|>` in any message to switch per-turn
- **Thinking state machine**: `ns_flags.enable_thinking` persists across all messages
- **`developer` role support**: Fully compatible with OpenAI API
- **Enhanced error messages**: All exceptions include diagnostic context
- **Tool call validation**: Name presence + arguments type + message ordering
- **`video_url` support**: Video inputs via `video_url` key detected

## Usage

### vLLM

```yaml
chat_template: /path/to/qwen3.6/chat_template.jinja
```

### llama.cpp / LM Studio

Place `chat_template.jinja` alongside your model files.

### Dynamic Thinking Toggle

```python
messages = [
    {"role": "system", "content": "You are helpful. <|think_off|>"},
    {"role": "user", "content": "What is 2+2?"},           # no thinking
    {"role": "assistant", "content": "4"},
    {"role": "user", "content": "<|think_on|> Prove Fermat's theorem"}  # thinking ON
]
```

### vLLM Template Parameters

```python
response = client.chat.completions.create(
    model="Qwen3.6-27B-FP8",
    messages=messages,
    extra_body={
        "chat_template_kwargs": {
            "enable_thinking": False,    # disable thinking
            "preserve_thinking": True     # keep thinking blocks in history
        }
    }
)
```

## Compatibility

| Engine | Status |
|--------|--------|
| vLLM >= 0.6 | Tested (primary) |
| llama.cpp | Compatible |
| LM Studio | Compatible |
| Ollama | Compatible |
| SGLang | Compatible |

## Related

- **GitHub**: [GeerMrc/qwen3-enhanced-chat-template](https://github.com/GeerMrc/qwen3-enhanced-chat-template) — full documentation (EN/CN), optimization details, client guide
- **Upstream**: [froggeric/Qwen-Fixed-Chat-Templates](https://modelscope.cn/models/froggeric/Qwen-Fixed-Chat-Templates)