OpenMobile-2 brings open-source mobile agents closer to closed models
Researchers from Nanjing University and SenseTime present OpenMobile-2, an open-source project of smartphone agents trained on over 110 simulated applications. Results improve on AndroidWorld, MobileWorld, and SPA-Bench, with a clear gain as app diversity increases.
OpenMobile-2, an open source smartphone agent project led by researchers from Nanjing University, SenseTime, and other universities, shows that a free model can now approach the best closed systems on mobile tasks. Published on October 25, 2026, the study describes two agents based on Qwen3.5 and Qwen3.6, with 9 billion and 27 billion parameters respectively, as well as the complete environment to reproduce the training. The challenge goes beyond a simple benchmark: on mobile, agents must understand shifting interfaces, persistent sessions, and applications that do not always offer APIs. And this is where the open source model begins to seriously overshadow others.
Training an agent on real apps remains difficult, so researchers rebuilt them
The team chose a pragmatic approach: rather than relying solely on real applications, which are difficult to reset and costly to handle, they rebuilt 35 popular apps, mostly in China, including Taobao, JD, DiDi, and Meituan, along with services like Slack and Google Drive. By combining these reconstructions with already simulated applications and apps running in an Android emulator, the training ground exceeds 110 applications. The simulations automatically reset to their initial state, while the agent only sees screenshots, without access to internal data. In other words, it learns almost like a human user, but without lukewarm coffee or forgotten passwords.
The researchers say they generated about 11,600 example workflows and more than 2,000 training tasks with automatic success verification. Google's Gemini 3.1 Pro served as the teacher model to produce the examples. The system thus remains open on the distribution side, but not entirely on the intellectual dependency side: OpenMobile-2 partly learns from a closed model.
The diversity of applications matters more than data volume
The most interesting result concerns a very concrete point: with a constant data budget, increasing the number of applications seen during training improves performance more than multiplying examples on just a few apps. In the study, the success rate on never-before-seen applications rises from 60.7% to 67.5% when application diversity increases. This logic also holds on benchmarks. On AndroidWorld, the large model improves from 67.1 to 79.9 points, and the small one from 54.3 to 74 points. On MobileWorld, more difficult because it covers long and multi-app tasks, the large model rises from 35.9 to 50.4 points.
The best improvement appears on real Android phones. On SPA-Bench, the 27-billion-parameter model goes from 31.9 to 59.6 points, the best score among open models according to the authors. On the same test, the OpenMobile-2-27B agent still lags behind Gemini 3.1 Pro. Reinforcement learning brings more modest gains, especially in environments already similar to pre-training ones; on AndroidWorld, the large model even slightly declines after this phase, from 80.8 to 79.9 points.
Interface shortcuts change the method, not yet the conclusion
The study also tests a hybrid approach: more than 50 simulated applications expose direct commands, like "add to cart," which allow avoiding certain menus. The agent can then combine classic gestures and application tools. Without specific training, it does not really use these shortcuts. Once accustomed, however, it solves up to 10.3 additional points of tasks while reducing the average path by about five steps. The authors note that this ability even transfers to tools not seen during training.
The project also relies on a custom benchmark of 215 tasks, nearly half covering multiple applications. Success is not judged by another language model, but by a program comparing data states before and after execution. The researchers say they verified there is little overlap between training and test tasks. In this context, OpenMobile-2 does not prove that an open agent has already caught up with the best closed systems everywhere. It mainly shows that in 2026, on mobile, the variety of environments matters as much as model size, and sometimes more.
Key points
- OpenMobile-2 covers more than 110 simulated applications.
- The large model reaches 59.6 on SPA-Bench, on real Android.
- 60.7% becomes 67.5% with more apps, at constant data.
- 35 apps were rebuilt by Nanjing University and SenseTime.
- Gemini 3.1 Pro served as the teacher model.
- Publication dated October 25, 2026.
By the numbers
- 110+ applications — training environment, according to the OpenMobile-2 study.
- 11,600 workflows — examples generated for learning.
- 2,000+ tasks — training tasks with automatic verification.
- 215 tasks — custom benchmark, nearly half multi-app.
- 59.6 points — score of the 27B model on SPA-Bench, on real Android devices.