OpenMobile-2: an open source smartphone agent, trained on over 110 applications, heels proprietary models

Researchers at the University of Nanjing and SenseTime publish an open mobile agent of 27 billion parameters that only divests in front of Gemini 3.1 Pro on a real-life test bench.

OpenMobile-2 is the result of a collaboration between Nanjing University, SenseTime and several other institutions. The team built a training environment of more than 110 simulated applications, including 35 reconstructions of applications that are widely used in China (Taobao, JD, DiDi, Meituan) or around the world (Slack, Google Drive). It resulted in about 11,600 example courses and more than 2,000 exercise tasks to teach models how to control a phone.

Two variants, with 9 and 27 billion parameters, are built on a Qwen architecture. The largest scores 79.9 points on AndroidWorld, 50.4 on MobileWorld, and 59.6 on SPA-Bench, which tests agents on real devices: the highest score among open models, with only Gemini 3.1 Pro doing better among the systems evaluated. The authors' main lesson is that the variety of apps matters more than the volume of data. For the same amount of data, going from 15 to 53 apps increased the success rate on never-before-seen apps from 60.7% to 67.5%. The environment, data, and reproduction instructions are published as open source.