OpenMobile-2 open-source smartphone agents rival proprietary models
openmobile-2 project trains ai smartphone agents with 110 simulated apps, achieving 59.6 points on spa-bench, the highest among open models. this shows app diversity matters more than data volume.
researchers from nanjing university, sensetime, and other institutions developed openmobile-2 to train ai smartphone agents in a simulated environment with over 110 apps. they released two models, built on qwen3.5 with 9 billion parameters and qwen3.6 with 27 billion parameters, along with the training environment and data. the larger model achieved 59.6 points on spa-bench with real android devices, the highest score among open models. the project shows that increasing app diversity during training improves performance more than sheer data volume. the researchers rebuilt 35 popular apps, including taobao and slack, to create a practice ground for the agents.