Agents · Updated 25 Oct, 12:25 pm IST
OpenMobile-2 shows open smartphone agents can rival proprietary models
Why it matters for readers: It demonstrates how training agents on diverse simulated apps helps them learn to operate real app interfaces.
- OpenMobile-2 consists of two agent models built on Qwen3.5 (9B) and Qwen3.6 (27B) and is released with code, data, and training environment.1
- The team rebuilt 35 popular apps (mainly Chinese) and combined them with simulations and emulator apps to produce practice data across more than 110 applications.1
- Researchers argue that diversity of apps used for training matters more than simply adding more data, and that agents learn primarily through practice.1
- Human reviewers iteratively refined the rebuilt apps so workflows and interfaces matched originals, enabling repeatable, resettable training without real-account problems.1
Get a brief like this every morning
Uzha reads hundreds of sources and gives you the stories that matter for your work, with every source linked. Free.
Get started