A team of authors from several institutions released MAI-Gym, an open‑source platform for building, testing and training AI agents that interact with graphical user interfaces (GUIs). The code, data and a 9‑billion‑parameter model checkpoint are available now for anyone to download. MAI-Gym works with mobile phones, desktop computers and web browsers, and can run up to 20,000 parallel environments.
What does MAI-Gym provide?
The platform includes a shared rollout pipeline that keeps agent behavior identical during evaluation and reinforcement‑learning (RL) training. It ships with ten established benchmark suites and 2,000 verified tasks covering apps such as ticket‑booking, email, and settings screens. MAI‑Gym also can synthesize new mock apps automatically, creating resettable environments and cross‑app worlds that are checked for correct execution.
Results with the released 9B model
Using the framework, the authors trained MAI‑Gym‑9B, a unified agent for both mobile and desktop use. On the MobileWorld benchmark the model scored 78.4 % success, and on OSWorld‑Verified it reached 76.2 %. In a transfer test, the agent learned a ticket‑booking task in a mocked app and then performed the same task on a real phone, raising success from 28.8 % to 48.1 %.
Why it matters
Developers can now train AI assistants that work across phones, PCs and browsers without building their own massive infrastructure. The open data and model let researchers test new ideas quickly, and the transfer result suggests that training in simulated apps can improve real‑world performance.
