Qwen-UI-Agent: A Foundation GUI Agent for Phones, Computers, and Browsers
A new preprint describes Qwen-UI-Agent, which mixes screen taps and typed commands across mobile, computer, and browser tasks.
What the researchers built
Qwen-UI-Agent is described in a technical report posted to arXiv on 30 Jul 2026, authored by Hanzhang Zhou and 15 co-authors, titled "Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents." The paper frames GUI agents as having the potential to become a general purpose executor over existing digital devices, and says its goal is agents that operate reliably on real devices, execute workflows across platforms, combine GUI interaction with command-line execution, complete long-horizon tasks, proactively initiate useful services, and autonomously improve their own capabilities with minimal human effort. Guided by that vision, the authors present Qwen-UI-Agent as a real-world centric foundation GUI agent spanning mobile, computer-use, web, and DeepSearch environments.
How the system is built
According to the preprint, Qwen-UI-Agent combines diverse sandbox environments with a large-scale real-device mobile runtime. Its unified action space interleaves GUI operations with command-line execution and generates batched actions in a single model turn. The authors describe an AutoResearch-style data flywheel that uses agents to construct tasks and environments, diagnose failures, and plan subsequent iterations. Training relies on online reinforcement learning across trajectories exceeding 100 turns, with more than 10,000 concurrent environments accelerating rollout, plus a lightweight harness layer that supports proactive service initiation and stateful workflows across mobile and computer.
What the researchers report measuring
The authors report state-of-the-art results on mobile-use benchmarks, plus competitive scores on computer- and browser-use tasks, naming frontier systems Opus 4.8, Gemini 3.1 Pro, and GPT-5.6 Sol as the comparison points. On mobile use, the paper reports 82.1% on MobileWorld, 92.2% on MobileWorld-Real, and 97.5% on AndroidDaily. On computer use, it reports 79.5% on OSWorld-Verified and a 40.0% partial-progress score on OSWorld-v2. On browser use and GUI grounding, the paper reports 73.6% on WebArena and 81.5% on ScreenSpot-Pro. These are the authors' own reported figures from a preprint that has not gone through peer review, so this describes what the paper claims rather than an independently confirmed result.
Key facts
- Qwen-UI-Agent is presented in a technical report submitted to arXiv on 30 Jul 2026 by Hanzhang Zhou and 15 co-authors.
- The system interleaves GUI operations with command-line execution and can generate batched actions in a single model turn.
- Training uses online reinforcement learning on trajectories exceeding 100 turns across more than 10,000 concurrent environments.
- The paper reports 82.1% on MobileWorld, 92.2% on MobileWorld-Real, and 97.5% on AndroidDaily for mobile-use tasks.
- On computer- and browser-use tasks, the paper reports 79.5% on OSWorld-Verified, a 40.0% partial-progress score on OSWorld-v2, and 73.6% on WebArena.
Why it matters
The paper is a preprint, meaning these results have not been independently verified or peer-reviewed, but the authors' own reported numbers describe a single system attempting mobile, computer, browser, and search tasks through one shared action space instead of separate tools for each surface. That framing, tapping and typing through one unified action space, is what teams building device-controlling agents will want to weigh against the reported scores.
Check your understanding
What score does the preprint report for Qwen-UI-Agent on MobileWorld-Real?
Which frontier models does the paper compare Qwen-UI-Agent against on computer- and browser-use tasks?