option
Home
Flash News
Content
BrianWalker
BrianWalker
August 21, 2026

Alibaba has released Qwen-UI-Agent, a GUI agent foundation model designed for real-world use. It functions across mobile, desktop, web, and deep search platforms, breaking limits of traditional simulation testing by operating directly on complex real devices. The model outperforms major international flagship models in multiple tests, achieving scores above 80% in mobile tests and first place in the WebArena web test. It also features a real-device benchmark with over 400 tasks and supports batch command execution to boost efficiency. The system includes a security mechanism that rejects illegal requests and pauses sensitive operations for user confirmation. It can undergo online reinforcement learning for long-term tasks through adaptive curriculum training. The project is available at the provided link.

Alibaba has released Qwen-UI-Agent, a GUI agent foundation model designed for real-world use. It functions across mobile, desktop, web, and deep search platforms, breaking limits of traditional simulation testing by operating directly on complex real devices. The model outperforms major international flagship models in multiple tests, achieving scores above 80% in mobile tests and first place in the WebArena web test. It also features a real-device benchmark with over 400 tasks and supports batch command execution to boost efficiency. The system includes a security mechanism that rejects illegal requests and pauses sensitive operations for user confirmation. It can undergo online reinforcement learning for long-term tasks through adaptive curriculum training. The project is available at the provided link. Alibaba has released Qwen-UI-Agent, a GUI agent foundation model designed for real-world use. It functions across mobile, desktop, web, and deep search platforms, breaking limits of traditional simulation testing by operating directly on complex real devices. The model outperforms major international flagship models in multiple tests, achieving scores above 80% in mobile tests and first place in the WebArena web test. It also features a real-device benchmark with over 400 tasks and supports batch command execution to boost efficiency. The system includes a security mechanism that rejects illegal requests and pauses sensitive operations for user confirmation. It can undergo online reinforcement learning for long-term tasks through adaptive curriculum training. The project is available at the provided link.
Comments (0)
0/300
OR