OpenMobile-2 proves open-source smartphone agents can rival proprietary models

Why it matters

Nanjing University and SenseTime, working with other universities, have released OpenMobile-2, an open-source project for training AI agents that operate smartphone apps. The team built a simulated practice environment covering more than a hundred applications, recreating popular services and pairing them with automated task checks. They released two agent models based on Qwen3.5 and Qwen3.6, along with the training environment, data, and reproduction instructions. The central finding is that app diversity in training matters more than data volume: holding data constant while increasing the number of practice apps improved success on unseen applications. On real Android devices, the larger model achieved the highest score among open models on SPA-Bench. The work addresses a structural problem: agents learn through practice, but such data is largely held by large companies, leaving open models behind. The project also explores direct commands as shortcuts, though the researchers describe that part as experimental.