Open-source phone agent OpenMobile-2 gains from training on more apps, not more data

Key points

OpenMobile-2 releases two open agent models, 9B (Qwen3.5) and 27B (Qwen3.6), plus training environment, data and reproduction instructions.
Training spans 110+ simulated apps; with data held constant, more apps lifted success on unseen apps from 60.7 to 67.5 percent.
The 27B model hit 59.6 on SPA-Bench on real Android devices, the best open-model score per the researchers, up from 31.9.
Training data came from Google's Gemini 3.1 Pro as teacher, so the open model learns from a closed system.

Why it matters for marketers: Phone agents that operate apps for users are getting more capable in open form, so teams planning app-based automation or mobile journeys should expect agents to navigate their interfaces.

Researchers from Nanjing University, SenseTime and several other universities released OpenMobile-2, an open-source project for training AI agents that operate smartphone apps. It includes two models built on Qwen3.5 (9 billion parameters) and Qwen3.6 (27 billion), along with the full training environment, data and instructions to reproduce the work.

The team rebuilt 35 apps as simulations, including Taobao, JD, DiDi, Meituan, Slack and Google Drive, and combined them with existing simulations and emulator apps to reach more than 110. Agents see only screenshots, as a person would. The key finding: with the same amount of data, training on more apps raised success on unseen apps from 60.7 to 67.5 percent.

On benchmarks, the larger model rose from 67.1 to 79.9 on AndroidWorld and from 35.9 to 50.4 on MobileWorld. On real Android devices it scored 59.6 on SPA-Bench, up from 31.9, which the researchers call the best among open models. Reinforcement learning added only limited gains. The training examples came from Google's Gemini 3.1 Pro, so the open model still draws on a closed one.

For marketers and content teams, the result suggests agents that shop, book and message inside apps will keep improving, and open versions lower the barrier for others to build them. Apps and mobile experiences may increasingly be used by software, not only people.

Watch the experimental direct commands, such as "add to cart." Trained agents solved up to 10.3 percentage points more tasks with about five fewer steps, but these interfaces exist only in the simulated apps for now.