Alibaba releases Qwen-UI-Agent: 27B model tops 5 of 6 GUI benchmarks, beating GPT-5.6 and Claude Opus 4.8
智东西 · wechat · 2026-08-20
Alibaba has released Qwen-UI-Agent, a GUI agent foundation model for mobile, desktop, web and deep-search environments that operates devices via virtual clicks and CLI execution. It takes first place in 5 of 6 core GUI benchmarks, beating GPT-5.6 Sol, Claude Opus 4.8 and ByteDance's Seed2.1 Pro — with the main version at just 27B parameters (built on Qwen3.5), plus 35B-A3B and 4B variants. It hits 92.2% success on the real-device benchmark MobileWorld-Real.
The key ingredient is a real-device training environment of 100+ physical phones and 150+ apps, with health-aware scheduling and virtual-screen tech boosting concurrent rollout efficiency 20x, directly attacking the sim-to-real gap. GUI, CLI and API calls share one unified action space with batched actions; training uses an SFT → ActionRL → Online RL data flywheel, improving success by 7%+, cutting reasoning tokens 21.3%, and reducing false-success rates 11.2%.
The paper also finds 52% of baseline failures come from real-world interference (pop-up ads, UI misreads) and 40% from execution deficits. For safety, the model refuses high-risk requests outright and pauses for confirmation on payments, data deletion and privacy actions. Technical report and code are public.
More from coding & agent
- Agents fail to reconsider strategy during post-training execution — omarsar0 · 2026-08-20
- Leaked System Prompt: Domestic Giant's Client Uses 3-Layer Memory & MCP Routing — vista8 · 2026-08-20
- Addy Osmani uses Agent loops to triage 90 daily PRs — rseroter · 2026-08-20
- One-man startup uses GPT-5.6 for cross-project agent collaboration — every · 2026-08-20
- ZenML releases Agent Trace Viewer to inspect coding agent trajectories locally — strickvl · 2026-08-20
- Claude Code + Opus 5 saturates ARC-AGI-3 via falsifiable predictions — scaling01 · 2026-08-20