OpenMobile-2 pushes open smartphone agents closer to closed rivals

Open-source smartphone agents have taken a meaningful step closer to the closed systems they are often compared with. OpenMobile-2, a research project from a large academic and industry team, combines a much broader training environment with openly released data, models and recipes.¹

Why the environment matters

The project’s MobileGym++ playground covers more than 110 applications. That includes 35 simulated commercial-style apps with persistent state, cross-app workflows and automatic resets, alongside apps running in Android emulators. The researchers say this makes it possible to train and test agents across varied tasks without the expense and mess of repeatedly resetting real accounts and devices.¹

The team released nearly 12,000 mobile interaction trajectories and more than 2,000 executable reinforcement-learning tasks with automatic verifiers. The training setup is designed to check whether an action really changed the app state, rather than accepting an agent’s claim that it succeeded.¹

The results

OpenMobile-2 comes in 9-billion- and 27-billion-parameter versions. The larger model scored 79.9 on AndroidWorld and 50.4 on MobileWorld, according to the paper. On real-device SPA-Bench testing, its score rose from 31.9 for the earlier setup to 59.6, which the researchers describe as the strongest result among open models.¹

One of the more useful findings is about variety rather than raw volume. Holding the amount of training data steady, the researchers increased the number of apps represented in the environment and saw success on unseen apps rise from 60.7% to 67.5%. In other words, an agent that practises across more kinds of software may generalise better than one fed more examples from a narrow set of apps.¹

OpenMobile-2 also tests a hybrid approach. More than 50 simulated apps expose direct, app-native commands alongside normal taps, swipes and typing. After targeted training, the agents solved up to 10.3 percentage points more tasks while using roughly five fewer steps. The authors stress that these interfaces remain experimental, and the tools are currently available inside the simulated environment rather than being a universal shortcut for every Android app.¹

The project’s official site makes the code, datasets, model checkpoints and training recipes available, while the arXiv paper provides the benchmark details and methodology.¹² The Decoder’s report provides additional context on how the results compare with other mobile-agent systems.³

Our opinion

OpenMobile-2 is interesting less because one benchmark number beats another and more because it treats the training environment as a core part of the model. Rebuilding a broad, resettable world is a practical answer to the data bottleneck facing open computer-use projects. The important caveat is that simulated commercial apps are still simulations: real-world reliability, privacy and permission handling remain harder problems. Even so, releasing the environment and the checks makes this a more useful contribution than another opaque leaderboard result.