OpenMobile-2 proves open-source smartphone agents can rival proprietary models
decoder ·
Researchers from Nanjing University and SenseTime have released OpenMobile-2, an open-source AI smartphone agent that rivals proprietary models in performance. The project utilizes two Qwen-based models with 9 billion and 27 billion parameters, trained on a simulated environment containing over 110 applications. A key finding of the study is that training diversity across many apps significantly improves success rates on unseen applications more than increasing data volume alone. On real Android devices, the larger model achieved a score of 59.6 on SPA-Bench, marking the highest result among open-source models. The training setup involved rebuilding popular Chinese and global apps to allow for safe, repeatable practice without login or purchase barriers. Additionally, the agents were trained to use direct commands when available, allowing them to skip complex menu navigation and complete tasks in fewer steps. While reinforcement learning provided some benefits, the primary gains came from the hybrid approach of combining screen actions with exposed app tools. This work demonstrates that open-source agents can effectively compete with closed systems like Google's Gemini through innovative training methodologies.
decoder