OpenMobile-2 proves open-source smartphone agents can rival proprietary models

OpenMobile-2 is an open-source AI agent designed to operate on Android smartphones. The research team behind it built a training environment spanning more than 110 applications, including 35 replicas of widely used apps, to give the model broad exposure to real-world mobile interfaces.

A central finding from the work is that smartphone agents do not necessarily improve by ingesting larger volumes of data. Instead, performance gains come from training across a wider variety of apps, which teaches the agent to generalize across different UI patterns and interaction styles.

The model was evaluated on real Android devices, not just simulated environments, which adds practical credibility to the results. By releasing the training environment and model openly, the researchers aim to let the community build and benchmark smartphone agents without relying on proprietary systems.

This work matters because on-device AI agents are an emerging frontier where open-source models have historically lagged behind closed, well-resourced alternatives. If open training environments can close that gap, it lowers the barrier for developers and researchers to experiment with mobile AI assistants.

The approach also highlights a broader lesson for agent training: diversity of task environments may be a more efficient lever than sheer data scale, a principle that could apply beyond smartphones to other embodied or interface-grounded AI systems.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.