OpenMobile-2 Open-Sources a Mobile Agent: 79.9 on AndroidWorld and 110 Apps

The OpenMobile-2 team has released a near-frontier mobile agent. It open-sourced not only the resulting weights but the entire foundation: the environment, the data and the training recipes. The agent scores 79.9 on AndroidWorld and 50.4 on MobileWorld, and on real devices in SPA-Bench the result rose from 31.9 to 59.6. The work rests on 35 simulated commercial apps, about 12 thousand trajectories for supervised fine-tuning and more than 2 thousand RL tasks with automatic verification.

A 110-App Sandbox

The centerpiece of the work is MobileGym++. It is a set of 35 realistic, commercial-style apps with rich functionality and scenarios that span several apps at once. The environment is also fully controllable: state can be reset, and task outcomes are checked programmatically. This is critical for reinforcement learning, because the reward has to be computed automatically and reproducibly, and live services with accounts and changing content can't provide that.

Apps configured in Android emulators are added on top of the simulators. Together they form a platform of more than 110 apps. This diversity works against overfitting to a narrow set of interfaces, which is where many mobile agents usually break.

Data and a GUI-Tool Hybrid

On this base, the authors collected almost 12 thousand interaction trajectories for SFT and, by their own account, the largest open collection of verifiable RL data for mobile agents: over 2 thousand executable tasks with an automatic reward.

The second idea is experimental. Apps expose some of their functions as app-native tools, while the interface remains available. Within a single task, the agent alternates between taps and swipes and tool calls: where calling a function is faster, it calls it, and where it needs the screen, it works through the GUI. MobileGym++ implements more than 300 carefully scoped tools across 50+ apps.

For evaluation, the team added MobileGym++ Bench: long, complex scenarios in commercial apps. The same set of tasks can be run in "GUI-only" mode and in hybrid mode, so the contribution of the tools is measured directly.

Key figures:

• AndroidWorld: 79.9
• MobileWorld: 50.4
• SPA-Bench on real devices: from 31.9 to 59.6
• Environment: 35 simulated apps, more than 110 in total with emulators
• SFT data: about 12 thousand trajectories
• RL data: more than 2 thousand tasks with an automatic reward
• Tools: 300+ across 50+ apps

What Teams Can Try

For research groups, the easiest path is to take the ready-made environment with verifiable tasks rather than build their own testbed from scratch: training and evaluation then become comparable with published results. For teams building mobile assistants, the hybrid benchmark is the most useful part: it shows how much accuracy and how many steps are gained by exposing app functions as tools, compared with working purely through the screen. In essence, this is the same logic as function calling in LLMs, transplanted to a mobile interface.

There are caveats. The authors themselves call the hybrid mode experimental. The jump on SPA-Bench from 31.9 to 59.6 looks strong, but the abstract does not say which baseline the comparison starts from, so check the full paper before drawing conclusions. The "near-frontier" assessment also comes from the authors themselves. It can be checked right away: everything is published on the project page at os-copilot.github.io/OpenMobile2-Home.