Alibaba Launches Qwen-UI-Agent for Multi-Platform GUI Automation
Alibaba has officially launched Qwen-UI-Agent, a GUI agent foundation model designed to interact with mobile, desktop, web, and deep search environments. The model achieves state-of-the-art performance on major benchmarks, including scoring 79.5% on OSWorld-Verified and 73.6% on WebArena. This release marks a significant step forward in GUI automation, enabling AI agents to seamlessly transition from simulated tasks to executing complex, multi-step actions across real-world devices. By outperforming existing flagship models, it paves the way for more practical and reliable autonomous digital assistants. Qwen-UI-Agent integrates safety guardrails that automatically block high-risk requests and prompt users for confirmation during sensitive operations like payments or data deletion. Additionally, it supports command-line execution, outputs batched actions to reduce execution steps, and utilizes online reinforcement learning across 10,000 concurrent environments.
## BACKGROUND
GUI agents are AI systems designed to interact with computer interfaces just like humans do, by viewing screens, clicking buttons, and typing text. To evaluate these agents, benchmarks like OSWorld test their ability to perform open-ended tasks across real operating systems, while WebArena measures their performance on web-based workflows. GUI grounding is the underlying capability of mapping textual instructions to specific pixel coordinates or elements on a screen.