OpenMobile-2 shows open-source smartphone agents improve most with broader app training

OpenMobile-2 focuses on a practical bottleneck for smartphone agents: open models often lack access to the app-use data held by large companies. The researchers found that, with the same amount of training data, increasing the number of apps improved success on unseen applications from 60.7 to 67.5 percent. The takeaway is direct: for mobile agents, broad exposure to different app interfaces and workflows can matter more than simply adding more examples from a small set of apps.

The project uses an open training environment covering more than 110 applications, including 35 replicas of popular apps. The article says the rebuilt apps include China-focused services such as Taobao, JD, DiDi, and Meituan, along with globally used services like Slack and Google Drive. The setup is designed to avoid hard-to-repeat problems in real apps, such as logins, purchases, and state resets, while still presenting the agent with screenshots rather than internal app data.

OpenMobile-2 includes two agent models built on Qwen3.5 with 9 billion parameters and Qwen3.6 with 27 billion parameters, along with the environment, data, and reproduction instructions. On AndroidWorld, the larger model rose from 67.1 to 79.9 points, while the smaller model rose from 54.3 to 74. On SPA-Bench, which uses real Android devices, the larger model improved from 31.9 to 59.6 points, which the researchers say is the best result among open models.

The study also tests a hybrid setup where agents can combine normal screen actions with direct app commands, such as adding an item to a cart. More than 50 simulated apps offer these shortcuts, but the models tend to ignore them without targeted training. After training, the agents solved up to 10.3 percentage points more tasks and needed about five fewer steps, though the article notes that these interfaces currently exist only in simulated apps and are described by the researchers as experimental.