Online reinforcement learning (RL) with verifiable rewards has driven rapid progress in mathematics, coding, and tool use, yet scaling this success to GUI agents remains difficult due to costly interactive environments and a shortage of realistic training tasks. We introduce MAI‑Gym, an open all‑in‑one platform that unifies GUI agent evaluation, data collection, and online RL training across mobile, computer, and web environments.
MAI‑Gym employs a shared rollout pipeline to ensure consistent agent behavior during evaluation and training, supports ten established benchmarks, and scales to 20,000 parallel environment instances. Beyond scalable infrastructure, the platform automatically synthesizes diverse GUI applications and challenging tasks, including resettable mocked apps, coherent cross‑app worlds, and execution‑validated task verifiers.
Using MAI‑Gym, we train MAI‑Gym‑9B, a unified 9‑billion‑parameter agent that works on both mobile and computer platforms, achieving state‑of‑the‑art performance among open GUI agents under 10 B parameters: 78.4% on MobileWorld and 76.2% on OSWorld‑Verified. Crucially, RL performed in a mocked train‑ticketing app transfers to its real‑world counterpart, raising task success on a physical phone from 28.8% to 48.1%.
We release the full framework, synthesis skills, mocked apps, 2,000 verified tasks, and the model checkpoint to support reproducible research and further advances in scalable GUI agent training.
Review