Researchers have introduced OpenMobile-2, an open-source training environment and model designed to improve how smartphone AI agents navigate mobile software. Rather than relying solely on expanding dataset volume, the project emphasizes training agents across a wide spectrum of software interfaces. The new environment incorporates more than 110 distinct applications, including 35 simulated versions of widely used software. According to the creators, evaluation on physical Android devices demonstrated that OpenMobile-2 reached top performance levels among open-source mobile agents.
Focusing on App Diversity
Smartphone AI agents operate by reading mobile graphical user interfaces and taking actions on behalf of users. While traditional machine learning scaling often focuses on feeding larger quantities of raw data into a model, OpenMobile-2 indicates that exposure to varied application workflows is a key driver of agent competence. By exercising models across dozens of distinct tools and replicated popular services, the framework aims to help open-source systems approach the task-handling capabilities typically reserved for closed, proprietary models.
Unanswered Questions and Next Steps
Initial details published by The Decoder leave several technical parameters undisclosed. The announcement did not specify the academic or corporate identities of the research team, nor did it name the underlying base architecture powering the OpenMobile-2 model. Additionally, the report did not detail the exact scoring metrics used during testing on physical Android devices, and it remains unclear when or where the training environment and code will be fully accessible to the public for independent testing.
As mobile artificial intelligence shifts toward autonomous task execution on personal handsets, open training environments provide an alternative path for community-driven development. Further testing by third-party researchers will clarify whether training across diverse apps can reliably close the gap with proprietary alternatives.
Lead image: Tima Miroshnichenko / Pexels
