OpenMobile‑2: Open‑Source Smartphone Agents Match Proprietary Models on 110‑App Benchmarks

Researchers from Nanjing University, SenseTime, and other institutions released two open‑source smartphone agents built on Qwen3.5 (9 B parameters) and Qwen3.6 (27 B parameters). They provided the full training environment, data, and reproducibility instructions. The team rebuilt 35 popular Chinese apps (Taobao, JD, DiDi, Meituan) and added globally used services (Slack, Google Drive), creating a practice set of over 110 applications. Using simulated and real Android environments, they generated ~11,600 example workflows and >2,000 practice tasks, with automated success checks. The agents were trained with the help of Google’s Gemini 3.1 Pro as a teacher model. On the AndroidWorld benchmark, the 27 B model rose from 67.1 to 79.9 points and the 9 B model from 54.3 to 74.0 points, surpassing the predecessor OpenMobile‑1 (64.7). On the MobileWorld benchmark, the 27 B model improved from 35.9 to 50.4 points. On real Android devices, the 27 B model achieved 59.6 points on SPA‑Bench, the highest score among open models and close to Gemini 3.1 Pro.