Qwen Launches Qwen-UI-Agent, a GUI Agent Foundation Model for Real-World Devices

On August 20, Qwen large model released Qwen-UI-Agent, positioned as a GUI agent foundation model for the real world, covering mobile, desktop, web, and DeepSearch environments. The model emphasizes reliable task execution on real devices, supporting mixed GUI and CLI operations, batch actions, long-horizon online reinforcement learning, and features safety boundary control and cross-platform proactive service capabilities.

In terms of performance, Qwen-UI-Agent achieves leading or near-leading results across multiple GUI and agent benchmarks, including MobileWorld, MobileWorld-Real, AndroidDaily, OSWorld-Verified, WebArena, and ScreenSpot-Pro. The project has also built a real-device environment covering over 100 real phones and more than 150 applications, and proposes an AI-driven data flywheel to accelerate processes such as task design, environment construction, training evaluation, and error analysis.

[TechFlow]