OpenMobile-2: open-source smartphone agents rival proprietary models
In brief
- OpenMobile-2 releases two open-source agent models (9B and 27B parameters) trained on over 110 apps.
- App diversity in training boosts performance more than extra data, per AndroidWorld tests.
- Multi-app task performance improves sharply, but reinforcement learning shows limited gains.
Researchers from Nanjing University, SenseTime, and other institutions released OpenMobile-2, an open-source family of smartphone agent models. They argue that training on a wider variety of apps improves performance more than simply increasing data volume.
The two models, based on Qwen3.5 (9 billion parameters) and Qwen3.6 (27 billion), were trained across over 110 apps, including 35 replicas of popular services like Taobao, JD, DiDi, and Meituan. The team generated 11,600 workflows and 2,000 training tasks with automated success checks. Results show significant gains on benchmarks such as AndroidWorld and MobileWorld, especially for multi-app tasks.
The authors emphasize that training environment diversity matters more than raw data size. They also note that reinforcement learning offers limited improvements, except in closely matching setups. The study explores direct commands to speed up interactions, but their effectiveness depends on targeted training.
This summary gives you the essentials. The full article, its figures and quotes are on The Decoder.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error