The OpenMobile-2 project releases two models for controlling phones and training resources
The OpenMobile-2 project is releasing models with 9 and 27 billion parameters, a training environment with more than 110 applications, data and a reproduction procedure. According to the authors, the larger model scored 59.6 points on SPA-Bench on real Android devices.
A team from Nanjing University, SenseTime and other universities is releasing two agent models for controlling phones as part of the OpenMobile-2 project. The smaller model is based on Qwen3.5 and has 9 billion parameters; the larger model is based on Qwen3.6 and has 27 billion parameters. The release includes training environments, data and instructions for reproducing the results.
The environment includes more than 110 applications: it combines simulations, 35 newly created replica applications and applications running in an Android emulator. The simulations can be reset automatically; the agent works with screenshots without access to internal application data. The team prepared approximately 11 600 example workflows and over 2 000 practice tasks with automatic checks of task completion. The examples were generated by Gemini 3.1 Pro from Google.
According to the authors, the larger model scored 59.6 points on SPA-Bench on real Android devices, which they describe as the best result among open models. With the same amount of training data, increasing the number of applications raised the success rate on previously unseen applications from 60.7 to 67.5 percent. These results come from measurements by the research team.
Why it matters
The released models and resources make it possible to reproduce research on phone control and explore how capabilities transfer between applications. The results from the team also provide a concrete reason to consider application diversity when assembling training data, alongside the number of examples.
Two audiences, two different impacts
What this means
For individuals
Developers get models and a documented procedure as a starting point for their own experiments with controlling phones using screenshots.
For a business
For companies developing mobile automation, the released environment provides a foundation for repeatable workflow testing with automatic checks of task completion.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.