Qwen-CUA – A Native Computer Use Agent from Alibaba Qwen and Others Operational

Qwen-CUA is a native Computer Use Agent jointly developed by the Qwen team and XLang Lab. Built on a 397B-A17B mixture-of-experts architecture, it perceives interface states solely through screenshots, without relying on the DOM tree or accessibility metadata, and directly controls browsers, desktop applications, and professional software through keyboard and mouse events. It scores 86.2 on the OSWorld-Verified benchmark, while the larger Qwen-CUA-Max has more than one trillion parameters and reaches 87.6.

Pure visual perception: Understands interface states solely from screenshots, without depending on HTML structure or accessibility metadata, providing stronger cross-platform versatility.
Native input simulation: Directly outputs keyboard and mouse events to perform actions in browsers, desktop applications, and professional software.
MoE architecture: Built on a 397B-A17B mixture-of-experts model, balancing performance and inference costs; the Max version has more than one trillion parameters for further performance gains.
Open technical report: The technical report and a reference demo are available on GitHub, making reproduction and further development easier for researchers.

Base architecture: Built on the Qwen3.5-397B-A17B mixture-of-experts architecture, with 397B total parameters and only 17B activated per forward pass, maintaining strong expressive power while controlling inference costs.
Hybrid attention mechanism: Alternates GatedDeltaNet linear attention and Gated Full Attention in a 75:25 ratio, reducing long-sequence computational complexity from O(n²) to near-linear and increasing throughput for 256K contexts by 19 times over traditional architectures.
Expert routing: Each layer contains 512 fine-grained routing experts. Top-10 selection and weight renormalization activate the experts, while 1 shared expert and sigmoid gating adaptively regulate the contribution of general knowledge to prevent information loss.
Pure visual perception: Perceives interface states solely through screenshots, with no dependence on HTML DOM trees, accessibility metadata, or operating system APIs, enabling cross-platform perception of arbitrary GUI applications.
Visual encoding and localization: Inherits the ViT architecture from the Qwen-VL series to support dynamic-resolution inputs, while MRoPE multimodal rotary position encoding jointly encodes temporal, height, and width information to precisely locate interface element coordinates.

Visit the repository: Search GitHub for xlang-ai/Qwen-CUA to open the project homepage.
Read the technical report: Download and read the project’s technical report to learn about the model architecture and training details.
Run the reference demo: Configure the environment according to the README and run the official reference demo to experience screenshot-driven automation.
Connect your own environment: Integrate the model with a local or cloud-based desktop/browser environment and execute tasks in a closed loop using screenshot inputs and keyboard and mouse outputs.

High benchmark scores: Scores 86.2 on OSWorld-Verified, while the Max version scores 87.6, placing it among the leading open-source CUA systems.
A truly native CUA: Unlike solutions that depend on APIs or combine multiple agents, Qwen-CUA is an end-to-end vision-action model with greater stability.
Cross-platform compatibility: It is not restricted by operating systems or application types and can control any software with a graphical interface.
Open-source ecosystem: Released under the Apache 2.0 license, with the technical report and demo publicly available for community extensions.

GitHub repository: https://github.com/xlang-ai/Qwen-CUA
Technical paper: https://github.com/xlang-ai/Qwen-CUA/blob/main/paper/Qwen-CUA.pdf

Automated software testing: Identifies interface elements from screenshots and performs clicks, text input, scrolling, and other actions, enabling cross-platform UI regression testing without writing specialized scripts.
RPA process automation: Replaces manual work for repetitive desktop tasks such as data entry, report generation, and ERP operations, without requiring changes to existing enterprise IT systems.
Accessibility assistance: Automatically identifies software interface elements and performs precise actions on behalf of visually impaired users or older adults, lowering the barrier to using professional software.
Scientific and professional software control: Automatically operates professional software such as MATLAB and SPSS without API interfaces, carrying out long-running tasks such as batch processing of experimental data and chart generation.
Cloud-based agent hosting: Deployed in cloud desktop environments to run background tasks 7×24 hours, including website inspections, data collection, and scheduled report downloads.