Alibaba has released Qwen-UI-Agent, an open-weight foundation model designed to operate graphical user interfaces on real phones, desktops, and browsers. The model tops several public leaderboards for mobile GUI use.
It scores 82.1% on MobileWorld. On the new MobileWorld-Real benchmark, run across more than 100 physical Android devices, it reaches 92.2%. On AndroidDaily it hits 97.5%.
On desktop it reaches 79.5% on OSWorld-Verified and 40.0% partial progress on the long-horizon OSWorld-v2 suite. On the web it posts 73.6% on WebArena.
The technical report, published on arXiv on July 30, 2026, describes a unified action space that interleaves GUI clicks, swipes, and text entry with command-line execution. The model can emit batched actions in a single turn, shortening execution trajectories.
Training combined sandbox simulation with a large-scale real-device runtime spanning over 100 phones and 150 applications. An AutoResearch-style data flywheel generates tasks, diagnoses failures, and plans subsequent iterations. Online reinforcement learning supports trajectories longer than 100 steps across more than 10,000 concurrent environments.
What's new
- Model: Qwen-UI-Agent (open-weight, Apache-2.0 license)
- Scope: Mobile (Android), desktop (Linux/Windows), web browsers, and cross-device workflows
- Key architectural features: Unified GUI + CLI action space, batched multi-action turns, proactive service initiation, stateful cross-device workflows
- Training infrastructure: 100+ real phones, 150+ apps, 10,000+ concurrent RL environments, trajectories >100 steps
- Benchmarks (selected): MobileWorld 82.1%, MobileWorld-Real 92.2%, AndroidDaily 97.5%, OSWorld-Verified 79.5%, OSWorld-v2 partial progress 40.0%, WebArena 73.6%, ScreenSpot-Pro (zoom) 81.5%
- Safety: Refuses illegal/high-risk requests; pauses for user confirmation on payments, data deletion, and privacy authorizations
- Availability: Technical report site at tongyi-mai.github.io/Qwen-UI-Agent/; GitHub Pages repo at github.com/Tongyi-MAI/Qwen-UI-Agent (website source only); model checkpoints and training code to be released via the MAI-UI organization
The project emphasizes closing the simulation-to-reality gap. MobileWorld-Real, a new benchmark introduced alongside the model, contains over 400 tasks executed on physical devices with varying network conditions, app versions, and UI states. Alibaba reports that Qwen-UI-Agent outperforms both closed-source models (GPT-5.6 Sol, Claude Opus 4.8, Gemini 3.1 Pro) and open-weight peers (Seed 2.1 Pro, Qwen 3.7 Plus) across every mobile benchmark tested.
On desktop, the agent trails Claude Opus 4.8 on OSWorld-Verified (79.5% vs. 83.4%). The report lists Opus 4.8 at 54.8% partial progress on OSWorld-v2, ahead of Qwen-UI-Agent's 40.0%. The browser result on WebArena (73.6%) edges out Opus 4.8 (71.9%) and GPT-5.5 (69.5%). Grounding scores on ScreenSpot-Pro reach 81.5% with zoom and 76.6% without.
Why it matters
GUI agents have largely been evaluated in sandboxed simulators. By training and benchmarking on a fleet of real phones, Alibaba addresses the distribution shift that causes models to fail when faced with dynamic content, network latency, and OS-level permission dialogs. The batched-action design and CLI interleaving also reduce step count, which matters for latency and cost when deploying agents at scale. The open-weight release lets researchers and developers audit the approach, fine-tune for vertical workflows, and integrate the agent into products without relying on proprietary APIs.
Our take
The real-device benchmark is the strongest signal: 92.2% on MobileWorld-Real suggests the sim-to-real transfer works better than prior public efforts. Desktop long-horizon performance (40% partial progress) remains the weak spot across the field, so Qwen-UI-Agent's lead there is modest. The safety gating — hard refusals on high-risk actions and confirmation pauses on payments — is a practical pattern product teams will likely copy.
