Researchers released 9B and 27B OpenMobile-2 smartphone agents with their training environments, data and reproduction recipes. A playground covering more than 110 apps, including 35 reconstructed apps, supplied about 11,600 demonstrations and over 2,000 verifiable tasks. The larger model improved from 31.9 to 59.6 on real-device SPA-Bench; increasing app diversity also improved success on unseen apps with data volume held constant. These are author-reported evaluations. Combining GUI actions with app-native commands was tested in simulated apps and does not guarantee automation across arbitrary real apps.