Alibaba Qwen-CUA: Free Open-Source Computer Use Agent with 397B Parameters
Alibaba Qwen-CUA: Free Open-Source Computer Use Agent with 397B Parameters
On August 5, 2026, Alibaba's Qwen team, in collaboration with XLang Lab, officially released Qwen-CUA, a native Computer Use Agent. This is China's first AI Agent model capable of perceiving UI states through screenshots and outputting keyboard/mouse events to control browsers, desktops, and professional software.
Core Technical Highlights
Qwen-CUA is built on a 397B-A17B Mixture-of-Experts (MoE) architecture. This means the model has 397 billion total parameters but only activates 17 billion parameters per inference, achieving an excellent balance between performance and efficiency.
Unlike traditional Computer Use approaches, Qwen-CUA does not rely on DOM trees or accessibility metadata. Instead, it perceives UI states directly through screenshots and outputs keyboard/mouse events. This approach offers several advantages:
- Broader applicability: Works with any application that has a screen display—web pages, desktop apps, or professional software
- Bypasses anti-bot detection: Doesn't interact with DOM, making it harder to identify as a bot
- Natural interaction: Simulates human visual perception, more intuitive for users
Performance Results
On the authoritative OSWorld-Verified benchmark, Qwen-CUA achieved an impressive 86.2 points. The trillion-parameter Max version reached 87.6 points, indicating the model can efficiently handle complex cross-application tasks.
OSWorld is a benchmark specifically designed to evaluate AI Agents' ability to execute tasks in real operating system environments, covering file management, software development, data visualization, web operations, and more. A score of 86.2 suggests Qwen-CUA has strong potential for large-scale practical application.
Free Open-Source & Usage
The best news is that Qwen-CUA is completely free and open-source. Users can:
- Download model weights from ModelScope or Hugging Face for local deployment
- Use Alibaba Cloud PAI platform for cloud inference
- Integrate into automation workflows via the Qwen-Agent framework
For regular users, we recommend using the cloud API approach—no local GPU configuration needed. For developers and enterprise users, downloading model weights for self-hosted deployment provides full data privacy control.
Application Scenarios
Qwen-CUA's main application scenarios include:
- Web automation: Auto-fill forms, scrape data, batch operations on web applications
- Desktop office work: Control Word, Excel, PPT for batch document processing
- Professional software: Operate CAD, Photoshop, Premiere for design tasks
- Cross-platform workflows: Seamless switching between applications for complex multi-step tasks
Comparison with Competitors
| Feature | Qwen-CUA | OpenAI Agent | Google Gemini Agent |
|---|---|---|---|
| Architecture | 397B-A17B (MoE) | Proprietary | Proprietary |
| Perception | Screenshot | Screenshot/DOM | Screenshot/DOM |
| Open Source | ✓ Free | ✗ Paid API | ✗ Paid API |
| OSWorld Score | 86.2 | Not disclosed | Not disclosed |
| Chinese Support | Native | Good | Good |
How to Use Qwen-CUA
Three recommended approaches:
Method 1: Cloud API (for regular users)
Use Alibaba Cloud PAI platform or call the official Qwen API directly. No GPU configuration needed, pay per call volume.
Method 2: Local Deployment (for technical users)
Download model weights locally and use vLLM or SGLang for inference. Requires at least 80GB VRAM A100/H100 GPU, or multiple consumer GPUs.
Method 3: Qwen-Agent Framework Integration
Use the official Qwen-Agent framework to quickly build automation agents, supporting tool calling, multi-turn dialogue, and task planning.
Conclusion
The release of Qwen-CUA marks a new milestone in China's AI Agent capabilities. With 397 billion parameters, an 86.2 OSWorld score, and completely free open-source licensing, it stands as one of the strongest Computer Use Agents available today. Whether for developers, enterprises, or individual users, Qwen-CUA can be leveraged to build powerful automation solutions.
For regular users, we recommend starting with the cloud API. For technically inclined users, local deployment offers more flexibility and privacy. As the model continues to iterate, Qwen-CUA is poised to become a benchmark product in the AI Agent space.