Researchers from Nanjing University and SenseTime have introduced OpenMobile-2, a set of open-source smartphone agent models. The project includes two versions based on Qwen architectures: a 9-billion-parameter model and a 27-billion-parameter model. The team released the full training environment, datasets, and instructions for reproduction.
To address the difficulty of training on real applications, the team rebuilt 35 popular apps, including Taobao, JD, and Slack. These simulations, combined with other environments, cover over 110 applications. The system allows for automatic state resets and uses Google's Gemini 3.1 Pro as a teacher model to generate 11,600 example workflows and 2,000 practice tasks.
The models demonstrate significant performance gains on benchmarks. On the AndroidWorld test, the larger model improved from 67.1 to 79.9 points, while the smaller version reached 74 points. On the SPA-Bench test for real Android devices, the 27-billion-parameter model increased its score from 31.9 to 59.6, which the researchers identify as the highest among open models.
The study highlights that training diversity is more critical than raw data volume. By increasing the number of apps used for practice while keeping the data amount constant, the success rate on unseen applications rose from 60.7 to 67.5 percent. Additionally, the researchers experimented with direct command shortcuts, which improved task completion by up to 10.3 percentage points.
