Researchers from Nanjing University, SI company SenseTime, and other institutions have released OpenMobile-2, a pair of open-source SI agents designed to operate smartphone applications. The release demonstrates that training on a diverse set of simulated environments can allow open models to rival proprietary systems, challenging the notion that data volume alone drives agent performance.
What Happened
The team released two SI models based on Qwen3.5 (9 billion parameters) and Qwen3.6 (27 billion parameters), accompanied by the full training environment, data, and reproduction instructions. To overcome the limitations of training on real apps—such as login barriers, irreversible purchases, and state persistence issues—the researchers rebuilt 35 popular apps, including Taobao, JD, DiDi, Meituan, Slack, and Google Drive, alongside existing simulations. This created a practice ground of more than 110 applications where agents could reset state and learn safely.
The models were trained using approximately 11,600 example workflows and over 2,000 practice tasks generated by Google's Gemini 3.1 Pro, which served as the teacher model. In a hybrid mode, the SI agents combine standard screen actions with direct function calls exposed by the apps. Human reviewers ensured the simulated apps matched the workflows and interfaces of the originals, while the agents interacted solely via screenshots, mimicking human usage.
Why It Matters
The study highlights that app diversity is more critical than sheer data volume for SI agent generalization. By holding the data amount constant and increasing the number of apps in the training set, the researchers saw the success rate on unseen apps rise from 60.7 to 67.5 percent. This suggests that the current bottleneck for open-source SI agents is not just the quantity of training data, but the breadth of environments they are exposed to during training.
Performance metrics support this approach. On the AndroidWorld benchmark, the larger OpenMobile-2 model scored 79.9 points, up from 67.1 for its predecessor, while the smaller model jumped from 54.3 to 74 points. On SPA-Bench, which tests performance on real Android devices, the larger model achieved 59.6 points, a significant increase from 31.9 points previously. The authors claim this is the best score among open models, trailing only Gemini 3.1 Pro. Reinforcement learning provided limited additional gains, sometimes even slightly decreasing scores on certain benchmarks, indicating that the initial supervised training on diverse simulations is the primary driver of capability.
The Bottom Line
OpenMobile-2 proves that open-source SI agents can compete with closed proprietary models when trained on a wide variety of simulated app environments. The release provides a reproducible framework for developing agentic SI tools that prioritize environmental diversity over raw data accumulation, potentially accelerating progress in smartphone automation and general SI research.