Skip to main content

Agent / GUI / Multimodal

Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

Hanzhang Zhou, Panrong Tong, Xu Zhang, Quyu Kong, Chenglin Cai, Tianyu Xia, Gongjie Zhang, Jianan Zhang, Long Li, Long Chen, Lei Wang, Gaole Dai, Pengxiang Li, Liangyu Chen, Yue Wang, Steven Hoi · Alibaba Tongyi Lab

arXiv:2607.28227264 upvotes

Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

Authors: Hanzhang Zhou, Panrong Tong, Xu Zhang, Quyu Kong, Chenglin Cai, Tianyu Xia, Gongjie Zhang, Jianan Zhang, Long Li, Long Chen, Lei Wang, Gaole Dai, Pengxiang Li, Liangyu Chen, Yue Wang, Steven Hoi

arXiv ID: 2607.28227

Problem: GUI agents could become the general-purpose executor over existing digital devices, but real-world use requires reliable operation on real devices, cross-platform workflows, interleaving GUI with CLI, long-horizon tasks, and proactive service initiation - capabilities that sandbox-trained agents do not have.

Key Methodology:

  • One agent spanning mobile, computer-use, web, and DeepSearch environments, combining diverse sandbox environments with a large-scale real-device mobile runtime
  • Unified action space that interleaves GUI operations with CLI execution, generating batched actions in a single model turn
  • AutoResearch-style data flywheel: agents construct tasks and environments, diagnose failures, and plan subsequent iterations
  • Online RL on trajectories exceeding 100 turns, with over 10,000 concurrent environments accelerating rollout
  • Lightweight harness layer supporting proactive service initiation and stateful workflows across mobile and computer

Key Results:

  • Mobile use: 82.1% on MobileWorld, 92.2% on MobileWorld-Real, 97.5% on AndroidDaily - state of the art on mobile-use benchmarks
  • Computer use: 79.5% on OSWorld-Verified and a 40.0% partial-progress score on OSWorld-v2
  • Browser use and GUI grounding: 73.6% on WebArena and 81.5% on ScreenSpot-Pro
  • Competitive against frontier models including Opus 4.8, Gemini 3.1 Pro, and GPT-5.6 Sol on computer- and browser-use tasks

Applied Context: The GUI-agent stack is consolidating: a single foundation model with a unified GUI+CLI action space, trained at real-device scale, now matches frontier models on computer and browser use while leading mobile. Builders of agent platforms should expect GUI operation to become a commodity model capability, and the interleaving of CLI and GUI actions points at where the next automation surfaces will come from.

Paper: arXiv:2607.28227