OpenMobile-2: an open phone agent trained on 110 apps scores 59.6 on real devices

A team from Nanjing University, SenseTime and other universities released OpenMobile-2, an open model and training recipe for agents that operate smartphone apps. The arXiv paper was posted on 9 October 2026, and The Decoder reported on it on 25 October. The authors say the agent is near-frontier and that its training environments, data and recipes are fully open. All figures below are the authors' own, or The Decoder's reading of the paper.

What was released

The paper describes MobileGym++, a set of 35 simulated commercial-style apps that can be reset and automatically checked, combined with apps in Android emulators to give a practice ground of more than 110 apps. The abstract lists nearly 12,000 interaction trajectories for supervised fine-tuning and more than 2,000 executable tasks with automatic rewards for reinforcement learning, which the authors call the largest open collection of its kind for mobile agents.

According to The Decoder, the two models are built on Qwen3.5 at 9 billion parameters and Qwen3.6 at 27 billion, and Google's Gemini 3.1 Pro produced the example workflows. That means the open model partly learns from a closed one. I did not check licence terms.

The results

The abstract reports 79.9 on AndroidWorld and 50.4 on MobileWorld, and says the model nearly doubles its score on SPA-Bench, a real-device test, from 31.9 to 59.6. The Decoder adds that the larger model went from 67.1 to 79.9 on AndroidWorld, that the predecessor OpenMobile-1 scored 64.7, and that 59.6 is the best open-model result on SPA-Bench, behind only Gemini 3.1 Pro. Reinforcement learning added little: on AndroidWorld the larger model scored 79.9 afterwards against 80.8 before, according to The Decoder.

Variety beat volume

The finding most useful outside the lab is an ablation. The Decoder reports that with the amount of training data held constant, increasing only the number of apps the agent practised on raised success on unseen apps from 60.7 to 67.5 percent. If you fine-tune or evaluate an agent for your own software, this suggests testing it across many different interfaces rather than piling up examples from a few.

Tool shortcuts, with a catch

In an experimental setting, more than 50 simulated apps expose selected functions as tools next to the screen, over 300 tools in all, so an agent can add an item to a cart instead of navigating menus. The Decoder reports that models mostly ignore these shortcuts until trained to use them; after training they solved up to 10.3 percentage points more tasks in about five fewer steps, and the habit carried over to tools not seen in training. The authors call the approach experimental, and the shortcuts exist only in the simulated apps.

What this does not show

• The rebuilt apps are mostly services popular in China, such as Taobao, JD, DiDi and Meituan, plus Slack and Google Drive. Nothing here tests an internal line-of-business app.
• The hybrid benchmark has 215 tasks and was built by the authors; The Decoder says success is checked by comparing app data before and after each run, not by a language model.
• Real-device results are on SPA-Bench, where the best system, per The Decoder, is still a closed model.

Treat OpenMobile-2 as a reproducible baseline for phone-agent experiments, not as evidence that an open agent can run unattended in a business app. Whether the open recipe holds up on enterprise software is unknown from these sources.