Can Open-Source AI Agents Finally Control Your Phone? Inside OpenMobile-2

According to The Decoder, the new open-source project OpenMobile-2 demonstrates that smartphone AI agents improve faster by training on a massive variety of applications rather than just crunching more raw data. Developed by researchers from Nanjing University and SenseTime, the project releases models based on Qwen with 9 billion and 27 billion parameters, aiming to close the gap with closed systems.

Why App Diversity Beats Raw Data Volume

Training AI models on real smartphone applications is notoriously difficult due to login walls, payment requirements, and the lack of automated state resets. To bypass this, the research team rebuilt 35 popular apps—such as Taobao, DiDi, Slack, and Google Drive—and combined them with existing simulations to create a practice environment covering over 110 applications.

By holding the data volume constant while increasing the number of training apps, the researchers pushed the success rate on unseen applications from 60.7% to 67.5%. On real Android devices, the larger 27-billion-parameter model achieved 59.6 points on SPA-Bench, which the authors highlight as the highest score among open models.

How to Access and Test OpenMobile-2

The research team has made the entire ecosystem accessible to the development community. Here are the core components released on GitHub and academic repositories:

Model Weights: Two agent versions built on Qwen architectures, featuring 9B and 27B parameters.
Training Environment: Complete simulation setups covering over 110 apps with automated state resets.
Datasets and Workflows: Approximately 11,600 example workflows and over 2,000 practice tasks generated using Google's Gemini 3.1 Pro as a teacher model.

Verdict: Is It Worth Your Time?

For developers and automation engineers working on mobile workflows, OpenMobile-2 represents a major leap toward independent local infrastructure. While experimental features like direct commands require targeted training to avoid being ignored by the models, the ability to run capable agents locally on Android hardware reduces reliance on costly proprietary APIs.