ComfyUI-Qwen 🐼

中文  ｜   English

Where Figma meets VSCode: Artistic vision meets engineering precision —— a romantic manifesto from designers to the code world.
✨ The Qwen3 ComfyUI Integration is a powerful tool designed specifically for ComfyUI workflows, aiming to seamlessly integrate Qwen series large language models (LLMs). This component provides both single-turn generation and multi-turn conversation nodes, supporting advanced features such as automatic model download, intelligent memory management, and a "thinking mode". ✨

---

🚀 Why Choose ComfyUI-Qwen?

At the intersection of creativity and technology, ComfyUI-Qwen serves as your intelligent creative companion. Through an intuitive visual interface, it seamlessly integrates Qwen series large language models, empowering creators, developers, and AI enthusiasts with powerful multimodal generation capabilities.

🌟 Key Features

Feature Category	Detailed Description
Model Support	Comprehensive coverage of Qwen3 series models (from 0.6B to 235B, catering to various scenarios)
Intelligent Device Adaptation	Automatically detects the optimal running device (GPU/CPU/MPS) and optimizes configuration
Dual-source Accelerated Download	Automatically selects the fastest download source (Hugging Face/ModelScope) with resume support
Quantization Technology	Supports 4-bit/8-bit quantization, significantly reducing VRAM requirements (min 2GB VRAM)
Thinking Mode	Enhances complex task processing capabilities with transparent reasoning and customizable thinking markers

💻 Installation

1. Navigate to the custom_nodes directory of ComfyUI
2. Clone this repository:

   git clone https://github.com/SXQBW/ComfyUI-Qwen3.git

3. Install dependencies:

   cd ComfyUI-Qwen3
   pip install -r requirements.txt

4. Restart ComfyUI

🎯 Examples

Multi-turn Conversation Node (Qwen Conversation)

This node supports full multi-turn conversations, maintaining conversation history, suitable for chatbot applications and prompt expansion in image generation workflows:

Prompt expansion result 1 for Flux image generation

Prompt expansion result 2 for Flux image generation

1. Select the model version (Qwen3-7B recommended by default)
2. Choose the quantization level (4-bit for low VRAM devices, 8-bit balances precision and performance)
3. Enable or disable "thinking mode" (suitable for complex reasoning tasks)
4. Enter your prompt text
5. Adjust the maximum generation length
6. Choose whether to unload the model after generation to free up resources

Single-turn Generation Node (Qwen Generation)

This node focuses on single text generation, suitable for prompt engineering, text expansion, and other tasks:

Single-turn generation node example 1

Single-turn generation node example 2

1. Select the model version
2. Choose the quantization level
3. Enable or disable "thinking mode"
4. Enter your prompt text
5. Adjust the maximum generation length
6. Choose whether to unload the model after generation to free up resources

🛠️ Technical Details

Memory Management

The component automatically detects your device (GPU/CPU/MPS) and selects the optimal running configuration:

• NVIDIA GPU users: Automatically uses CUDA and adjusts quantization based on VRAM size
• Apple Silicon users: Automatically uses MPS acceleration
• Low memory devices: Automatically downgrades to 4-bit quantization to save resources

Model Download

The component automatically tests download speeds from Hugging Face and ModelScope, selecting the fastest source. If download fails, it automatically tries the other source, with up to 3 retries.

Thinking Mode

The "thinking mode" is implemented through a special tagging mechanism, adding thinking process markers to the output during generation:

• When enabled: Shows the complete thinking process
• When disabled: Automatically filters thinking content, showing only the final result

📚 Supported Models

The following Qwen model versions are currently supported:

Model Name	Parameters	Recommended VRAM	Quantization Support
Qwen3-0.6B-FP8	600M	2GB	✅
Qwen3-0.6B-Base	600M	2GB	✅
Qwen3-0.6B	600M	2GB	✅
Qwen3-1.7B-FP8	1.7B	4GB	✅
Qwen3-1.7B-Base	1.7B	4GB	✅
Qwen3-1.7B	1.7B	4GB	✅
Qwen3-4B-FP8	4B	8GB	✅
Qwen3-4B-Base	4B	8GB	✅
Qwen3-4B	4B	8GB	✅
Qwen3-8B-FP8	8B	12GB	✅
Qwen3-8B-Base	8B	12GB	✅
Qwen3-8B	8B	12GB	✅
Qwen3-14B-FP8	14B	20GB	✅
Qwen3-14B-AWQ	14B	8GB	✅
Qwen3-14B-Base	14B	20GB	✅
Qwen3-14B	14B	20GB	✅
Qwen3-14B-GGUF	14B	8GB	✅
Qwen3-30B-A3B-FP8	30B	40GB	✅
Qwen3-30B-A3B	30B	40GB	✅
Qwen3-30B-A3B-Base	30B	40GB	✅
Qwen3-32B-FP8	32B	40GB	✅
Qwen3-32B-AWQ	32B	12GB	✅
Qwen3-32B	32B	40GB	✅
Qwen3-32B-GGUF	32B	12GB	✅
Qwen3-235B-A22B-FP8	235B	300GB	✅
Qwen3-235B-A22B	235B	300GB	✅

🤝 Contributing

We welcome community contributions! If you find issues or have suggestions for improvement, please submit an issue or pull request.

💌 Acknowledgments

Special thanks to the Qwen team for developing these powerful models, and to the ComfyUI community for their support!

The star you're about to click ✨
Is not just a gesture of approval, but a cosmic explosion where design thinking meets the code universe. When an artist's aesthetic obsession collides with a programmer's geek spirit – this might just be the most romantic chemical reaction on GitHub.

Click to Star and Witness the Cross-Disciplinary Revolution