Alibaba releases Qwen-UI-Agent mobile benchmark surpassing GPT 5

Alibaba releases Qwen-UI-Agent mobile benchmark that surpasses GPT 5.6 and Claude Opus 4.8. On August 20, Alibaba officially launched Qwen-UI-Agent, a GUI agent base model centered on the real world, covering mobile, computer, web and deep search (DeepSearch) environments. According to the official introduction, Qwen-UI-Agent has fully benchmarked and surpassed the industry's flagship model in multiple GUI benchmark tests. On the mobile side, it reached 82.1% on the MobileWorld benchmark, leading GPT-5.6 Sol and Claude Opus 4.8 by 12.0 and 14.6 percentage points respectively; it reached 92.2% on the real machine benchmark MobileWorld-Real, surpassing Gemini 3.1 Pro and Claude Opus. 4.8, GPT-5.6 Sol and other models; on AndroidDaily it is as high as 97.5%, close to the perfect score. On the computer side, it reached 79.5% on OSWorld-Verified, surpassing models such as GPT-5.5 and Gemini 3.1 Pro. In terms of browser and DeepSearch, it reached 73.6% on WebArena, ranking first among all compared models. In terms of GUI Grounding, it reached 81.5% on ScreenSpot-Pro, and also refreshed SOTA in the remaining four evaluation benchmarks. Qwen-UI-Agent has built a real-machine mobile environment covering more than 100 real mobile phones and more than 150 applications for task construction, trajectory collection, model training and evaluation, and built its own MobileWorld-Real real-machine benchmark (more than 400 tasks and more than 100 applications) to get through the last mile of "from simulation to reality". In addition to GUI operations, the model can also directly perform command line operations and output multiple actions in batches in a single decision. CLI actions account for nearly half of computer tasks, and about 40% of actions are output in batches. In terms of security mechanism, Qwen-UI-Agent will make security judgments throughout the entire task execution process: it will directly reject and terminate the task in the face of illegal or high-risk requests; in the face of sensitive scenarios such as payment, data deletion, and privacy authorization, it will proactively stop at key steps and explain the situation to the user. The model also supports online reinforcement learning training on ultra-long trajectories of more than 100 steps, and can be rolled out simultaneously with about 10,000 concurrent environments to continuously conquer long-distance tasks. via cnBeta.COM - Chinese industry information website