Researchers from Nanjing University, SenseTime and other institutions have released OpenMobile-2, an open training system for AI agents that operate smartphone apps. The package includes two Qwen-based models, data and instructions intended to make the results reproducible.
The environment covers more than 110 applications, including 35 replicas of services such as Taobao, JD, DiDi, Meituan, Slack and Google Drive. Simulated apps avoid logins, real purchases and persistent state, allowing researchers to reset tasks and collect repeatable training attempts more easily than on live services.
The central result is that application diversity mattered more than simply increasing data volume. With the amount of training data held constant, expanding the number of apps raised success on unseen applications from 60.7 to 67.5 percent. On real Android devices, the larger 27-billion-parameter model scored 59.6 on SPA-Bench, which the authors say is the best result among open models.
OpenMobile-2 can combine normal screen actions with functions exposed as tools. That may improve reliability where an app offers structured controls, while preserving visual interaction elsewhere. The scores still show substantial room for failure, and simulated replicas cannot capture every interface change, account restriction or safety consequence in production. The release is most useful as shared infrastructure for studying transfer across apps, not evidence that phone agents are ready for unattended access to purchases or private accounts.