Alibaba officially released the new GUI agent foundation model Qwen-UI-Agent on August 20. This model comprehensively covers mobile, desktop, web, and deep search (DeepSearch) environments, and has fully matched and surpassed several flagship models in the industry in multiple core graphical user interface benchmarks.
In specific benchmark performance, Qwen-UI-Agent demonstrates strong capabilities on mobile devices. It achieved a score of 82.1% on the MobileWorld benchmark, leading GPT-5.6Sol and Claude Opus4.8 by 12.0 and 14.6 percentage points respectively; on the real environment benchmark MobileWorld-Real, it achieved 92.2%, surpassing Gemini3.1Pro, Claude Opus4.8, and GPT-5.6Sol; on the AndroidDaily benchmark, it reached as high as 97.5%, approaching a perfect score. On desktop and browser platforms, it achieved 79.5% on OSWorld-Verified, surpassing GPT-5.5 and Gemini3.1Pro; on WebArena, it ranked first among all compared models with a score of 73.6%. In addition, it scored 81.5% on the GUI Grounding ScreenSpot-Pro test and set new SOTA records on the other four evaluation benchmarks.
To tackle the challenge of moving from simulation to reality, Qwen-UI-Agent built a real mobile environment covering more than 100 real smartphones and over 150 applications for task building, trajectory collection, model training, and evaluation. It also established its own MobileWorld-Real real-device benchmark, which includes over 400 tasks and more than 100 applications. In addition to conventional GUI click operations, the model supports direct command-line operations and batch outputs multiple actions in a single decision. Approximately 40% of the actions in desktop tasks are output in batch form.
In terms of security and long-term task processing, Qwen-UI-Agent integrates security mechanisms throughout the entire task execution process. When facing illegal or high-risk requests, the model will directly refuse and terminate the task; when dealing with sensitive scenarios such as payments, data deletion, and privacy authorization, it will proactively stop at key steps and inform the user. At the same time, the model supports online reinforcement learning training on ultra-long trajectories exceeding 100 steps, combined with about 10,000 concurrent environments simultaneously rolling out, continuously tackling challenging long-term tasks.
