Alibaba's Qwen team on August 20, 2026 released Qwen-UI-Agent, a GUI-based agent foundation model that operates real devices — phones, desktops, browsers and deep-search environments — by reading the screen and issuing clicks, keystrokes and swipes rather than calling APIs. The company says the release marks its first agent trained and evaluated end-to-end on live hardware.
Trained on live screens, not sandboxes
Where most agent benchmarks run inside synthetic sandboxes, Alibaba trained Qwen-UI-Agent by executing real workflows on real operating systems. About 40% of desktop tasks combine graphical actions with batched command-line output, letting the model chain multi-step operations in a single decision. The model is positioned as a base runtime other developers can layer product logic on top of.
Benchmark leadership across mobile, desktop and web
Qwen-UI-Agent scored 82.1% on the MobileWorld benchmark, beating OpenAI's GPT-5.6 Sol by 12.0 points and Anthropic's Claude Opus 4.8 by 14.6 points. It also posted 79.5% on OSWorld-Verified for desktop and took the top spot on WebArena for browser tasks. Alibaba says the results extend its lead in the agentic-AI race that began with July's Qwen 3.8-Max launch.
Safety guardrails for money and data
Because the agent physically drives the same interfaces a human uses, Alibaba built explicit stop-and-confirm behaviour into the runtime. The model refuses tasks flagged as illegal or high-risk and pauses before payments, privacy authorisations or destructive data operations, prompting the user with a clear description of the pending action. The company is also opening the platform to third-party developers, extending the ecosystem push behind the recent Qwen3.8-27B open-weights release.
Reporting based on coverage from Pandaily, Alibaba Cloud and Bloomberg.