Alibaba Launches Qwen-UI-Agent GUI Foundation Model, Scores 82.1% on MobileWorld

On August 20, Alibaba officially launched Qwen-UI-Agent, a real-world-centric GUI foundation model covering mobile, desktop, web and deep-search environments. The model achieved corresponding scores on multiple GUI benchmarks.

On the mobile benchmark MobileWorld, Qwen-UI-Agent reached 82.1%, leading GPT-5.6 Sol by 12.0 percentage points and Claude Opus 4.8 by 14.6 percentage points. On the real-device benchmark MobileWorld-Real, it reached 92.2%, surpassing Gemini 3.1 Pro, Claude Opus 4.8 and GPT-5.6 Sol. On AndroidDaily, it reached 97.5%.

For desktop, it reached 79.5% on OSWorld-Verified, surpassing GPT-5.5 and Gemini 3.1 Pro. For browser and deep search, it reached 73.6% on WebArena, ranking first among all compared models. For GUI grounding, it reached 81.5% on ScreenSpot-Pro, and also refreshed state-of-the-art results on the other four evaluation benchmarks.

Qwen-UI-Agent has built a real-device mobile environment covering more than 100 real phones and more than 150 applications, used for task construction, trajectory collection, model training and evaluation. It also self-built the MobileWorld-Real real-device benchmark, containing more than 400 tasks and more than 100 applications. The model can directly execute command-line operations and output multiple actions in batch in a single decision. In computer tasks, CLI actions account for nearly half, with about 40% of actions output in batch form. Regarding safety mechanisms, the model integrates safety judgment throughout the entire task execution process. For illegal or high-risk requests, it directly refuses and terminates the task. For sensitive scenarios such as payments, data deletion and privacy authorization, it proactively stops at key steps and explains the situation to the user. The model also supports online reinforcement learning training on ultra-long trajectories exceeding 100 steps, with about 10,000 concurrent environments for rollout.