On August 20, Alibaba's Qwen team released Qwen-UI-Agent, a foundation model for GUI agents built for real devices. The idea is simple to state: get the model to actually see a screen, then tap, swipe and type its way through a task from start to finish, the way a person would. Coverage spans phones, computers, the web, and a DeepSearch environment with live retrieval.
The scorecard Alibaba published is dense. On mobile, Qwen-UI-Agent scores 82.1% on MobileWorld, 92.2% on the real-device MobileWorld-Real, and 97.5% on AndroidDaily. On desktop, it hits 79.5% on OSWorld-Verified, 73.6% on the browser-task benchmark WebArena, and 75.0% on the Chinese retrieval benchmark BrowseComp-ZH. For locating interface elements, it reports 81.5% on ScreenSpot-Pro. On a subset of OSWorld-v2 tasks, Alibaba says it hits a 40.0% completion rate while cutting the number of execution steps by roughly 58%. The comparison set named alongside these numbers is GPT-5.6 Sol, Claude Opus 4.8 and Gemini 3.1 Pro.
A training ground built from 100-plus phones
What says more about where the investment went than the scores is the environment behind them. Alibaba built a lab with more than 100 real phones covering 150-plus apps, backed by a real-device benchmark of 400-plus tasks, capable of running 10,000 concurrent environments, with reinforcement-learning trajectories supported out to 100-plus steps.
This is an asset-heavy approach. GUI agents have mostly trained in simulators or web sandboxes in recent years — cheap and reproducible, but with a gap separating them from real hardware: real phones pop up ads, throw permission dialogs, and stall on loading screens when the network stutters, and those are exactly the moments where long tasks tend to fail. Bringing an actual device farm in-house trades hardware cost for distribution alignment.
That's also where the gap between single-step accuracy and long-task success shows up. Do the rough math: if the per-step success rate is 81.5%, the odds of a 100-step task completing in one clean run are close to negligible. What keeps a long workflow alive is the ability to notice a mistake, back out, and try another path — and real-device environments are precisely what generates that kind of failure sample, something a simulator struggles to produce naturally.
On desktop, GUI clicks mix with the command line
One detail on the desktop side: Alibaba says the system supports a mix of interface actions and command-line calls, with roughly half of desktop tasks involving a CLI invocation.
That's effectively an admission, in engineering terms. Trying to finish a task like "sort these files by date and zip them up" purely by clicking would balloon the step count out of control; an agent that can type a command line skips dozens of clicks outright. The trade-off is that the permission boundary gets more sensitive — the blast radius of one command is far larger than one misclick. Relaying work across devices is handled by a Harness framework, so a job left unfinished on a phone can be picked up and continued on a computer.
Where the safety line is drawn
"Faced with an illegal or high-risk request, it won't perform any interface action at all — it refuses outright and terminates the task," is how Alibaba describes the model's refusal policy. Sensitive actions such as payments, data deletion or changing authorizations require user confirmation first.
For users in China, how much that constraint actually holds up won't be clear until the model reaches real devices. High-risk operations on a phone often don't look obviously different from ordinary ones — a password-free payment confirmation screen and a plain "Next" button can look nearly identical pixel by pixel, and the model's judgment depends on context rather than how a control looks.
Who picks this up
The technical report is already up on arXiv, and the code repository sits on GitHub under Tongyi-MAI/MAI-UI. That repository previously open-sourced the MAI-UI series in four sizes from 2B to 235B parameters, under the Apache 2.0 license. How openly Qwen-UI-Agent's weights will be released hasn't been specified yet.
Judging by the likely rollout order, phone makers and PC makers stand to benefit first — they hold system-level permissions and already have a natural entry point for slotting an agent into a voice assistant. Third-party apps that want to plug in can't avoid the accessibility-permission hurdle, and that hurdle has stayed tightly regulated in China.
Sources: Qwen's official WeChat account, IT Home, CocoLoop, Kuai Technology; benchmark figures and real-device environment scale cross-checked against the Qwen-UI-Agent technical report and project page.