- Published on
OpenMobile-2 proves open-source smartphone agents can rival proprietary models
- Authors
- Name
- DeepHorus
Researchers from Nanjing University, SenseTime and other universities have released OpenMobile-2, an open-source project for training AI smartphone agents. It includes two Qwen-based models, a full training environment, data and reproduction instructions. The project aims to close a gap: most high-quality agent training data remains locked inside large companies.
OpenMobile-2 trains agents in a simulation covering more than 110 apps. The team rebuilt 35 popular services, including Taobao, JD, DiDi, Meituan, Slack and Google Drive, matching their workflows and interfaces. Each simulation resets automatically, and the agent sees only screenshots, mimicking human phone use.
The setup produced about 11,600 example workflows and over 2,000 practice tasks with automated checks. Google’s Gemini 3.1 Pro was the teacher model, meaning the open project learns from a closed system. The models use Qwen3.5 with 9 billion parameters and Qwen3.6 with 27 billion parameters.
On AndroidWorld, the larger model rose from 67.1 to 79.9 points and the smaller from 54.3 to 74; predecessor OpenMobile-1 scored 64.7. On MobileWorld, the larger model improved from 35.9 to 50.4. On real Android devices, it jumped from 31.9 to 59.6 points on SPA-Bench, which researchers call the best open-model result.
App diversity mattered more than data volume: with the same data budget, adding apps raised success on unseen applications from 60.7 to 67.5 percent. Reinforcement learning brought only limited gains; on AndroidWorld the larger model fell slightly to 79.9 after scoring 80.8. An experimental direct-command feature in more than 50 simulated apps offered little benefit because models ignored shortcuts.
The authors position OpenMobile-2 near the top among current smartphone agents, but acknowledge that real-device performance and shortcut use remain areas for improvement. The work shows open-source agents can narrow the proprietary gap. It also suggests training environments, not raw data volume, will shape future progress.
