OpenMobile-2 proves open-source smartphone agents can rival proprietary models
Summary
- Researchers from Nanjing University and SenseTime released OpenMobile-2, a set of open-source AI smartphone agent models built on Qwen architectures.
- The project uses a simulation environment covering over 110 apps to demonstrate that training app diversity matters more than sheer data volume.
- OpenMobile-2 includes 9B and 27B parameter models, along with the full training environment, data, and reproduction instructions.
- The 27B model achieved 59.6 points on SPA-Bench on real Android devices, marking the highest score among open-source models.
- The models also support experimental direct app commands, allowing agents to skip menu navigation and solve tasks in fewer steps.
Why it matters
OpenMobile-2 demonstrates that open-source smartphone agents can approach proprietary performance by leveraging diverse simulated app environments rather than massive proprietary training datasets. This provides researchers and developers with reproducible tools to advance mobile automation.