Smartphone agents learn to complete tasks by interacting with apps, but training on real services can be difficult to repeat. OpenMobile-2, an open-source project from researchers at Nanjing University, SenseTime, and other universities, offers a broad simulated environment for that practice. Its results suggest that exposure to more kinds of apps can matter more than simply adding examples from a narrow set.
A practice environment spanning more than 110 apps
Using real apps for training can involve logins, purchases, and app states that are hard to reset. To make practice repeatable, the researchers rebuilt 35 apps, including China-focused services such as Taobao, JD, DiDi, and Meituan, as well as Slack and Google Drive. They combined those with existing simulations and apps running in an Android emulator, creating a training environment with more than 110 apps.
Human reviewers compared the recreated apps with their originals and refined them to match their interfaces and workflows. The simulations can preserve details such as orders, budgets, and travel plans across sessions, and information can move between apps through the clipboard. At the same time, each simulation can return to its starting state automatically.
The agent interacts through screenshots, without access to the apps’ internal data. That setup is intended to resemble ordinary use, where a person sees what is on screen and acts through the interface. From this environment, the team produced about 11,600 example workflows and more than 2,000 practice tasks with automated checks for success.
More varied practice improved results on unfamiliar apps
The project releases two models: one built on Qwen3.5 with 9 billion parameters and another on Qwen3.6 with 27 billion parameters. The researchers also provide the training environment, data, and instructions for reproducing the work.
In a test that held the amount of training data constant, the researchers increased the number of apps used for practice. Success on unseen applications rose from 60.7 to 67.5 percent. The result points to a practical distinction: a fixed training budget can expose an agent to more interface patterns, or provide more examples from fewer apps.
The examples were generated by Google's Gemini 3.1 Pro, a closed model. So while OpenMobile-2 is released as an open project, its training process includes examples produced by a proprietary system.
Benchmark gains, especially on real Android devices
The models improved on established tests of smartphone interaction. On AndroidWorld, the larger model scored 79.9 points, compared with 67.1 before the reported improvement. The smaller model reached 74, up from 54.3. OpenMobile-1 had scored 64.7 on the same test.
On MobileWorld, which tests longer tasks involving multiple apps, the larger model's score increased from 35.9 to 50.4 points. The largest reported change was on real Android devices: the larger model rose from 31.9 to 59.6 points on SPA-Bench. The researchers describe that as the highest score among open models.
The team also built a benchmark with 215 tasks, nearly half of them involving multiple apps. Rather than asking a language model to judge whether a task was complete, a program compared stored app data before and after each attempt. The researchers checked training and test tasks for overlap and reported only minor similarities.
Shortcuts can help, but agents need practice using them
Some simulated apps offer direct commands, such as “add to cart,” that can bypass a sequence of screen taps. More than 50 apps include these commands, and agents can switch between direct commands and ordinary interface navigation.
Without specific training, however, the models often ignore the shortcuts. After training to use them, agents solved up to 10.3 percentage points more tasks while taking about five fewer steps. The researchers say this skill also transferred to tools the models had not encountered during training.
The shortcut approach remains experimental, and these direct command interfaces currently exist only in the simulated apps. Reinforcement learning produced limited additional gains, mainly in environments similar to training. On AndroidWorld, the larger model's score after that phase was 79.9 instead of 80.8 points.
OpenMobile-2 therefore offers evidence that broad, repeatable app practice can improve smartphone agents, including on unfamiliar apps and real devices. The results also show the limits of the approach: shortcuts need to be taught, reinforcement learning did not improve every score, and the examples came from a closed teacher model.