OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools

OpenMobile-2. Building Versatile Mobile Agents with Scalable Environments and App-Native Tools. By Chenyang Yan and twenty-three others, from Nanjing University, SenseTime, and four other institutions. Published on the ninth of October. Its code repository is worked out: about five hundred source files, with a readme and a licence.

OpenMobile-2 is described as a near-frontier mobile agent, but what sets the release apart is everything around the agent. Its training environments are fully open, and so are its recipes. Alongside the model come the apps it was trained in, the trajectories and tasks it learned from, a new tool-use setting, and a benchmark. The authors present this as three advances, with the benchmark added on.

Taken as an inventory, the release has four parts. An environment, MobileGym plus plus, with thirty-five commercial-style apps of its own, inside a playground of more than one hundred and ten apps in all. Nearly twelve thousand mobile interaction trajectories for supervised fine-tuning. More than two thousand executable tasks with automatic rewards for reinforcement learning, which the authors call the largest open collection of verifiable RL training data for mobile agents. And more than three hundred carefully scoped app-native tools, spread across more than fifty apps.

Each part is built for control. The thirty-five apps in MobileGym plus plus are simulated commercial apps: realistic, functionally rich, linked by cross-app workflows, yet fully controllable, so any of them can be reset and its results verified. The rest of the playground comes from newly configured apps running in Android emulators. The trajectories and tasks are curated on top of that foundation. The tools come from an experimental mobile-use setting in which apps expose selected functions as tools next to their ordinary graphical interfaces, so an agent can interleave GUI actions and tool calls within one task. To evaluate all this, MobileGym plus plus Bench poses complex, long-horizon commercial scenarios, and runs one shared task suite in two modes: GUI only, and hybrid GUI with tools.

The agent trained on it performs competitively on established benchmarks: a score of seventy-nine point nine on AndroidWorld, and fifty point four on MobileWorld. On real-device mobile use it shows promising transfer, nearly doubling performance on SPA-Bench from thirty-one point nine to fifty-nine point six. For anyone building mobile agents, the offer is the open stack behind those numbers: apps that can be reset and checked, trajectories to fine-tune on, verifiable tasks for reinforcement learning, and a shared suite for asking whether tools help.

Weights on their own would leave out most of what it takes to train an agent like this. Here the environments, the data, the tools and the evaluation are all on the table, and every resource is linked from the project's home page.