OpenMobile-2 open-source smartphone agents released, scoring 59.6 on SPA-Bench real-device test, best among open models

Researchers from Nanjing University, SenseTime and other institutions released OpenMobile-2, two open smartphone-agent models (9B built on Qwen3.5, 27B built on Qwen3.6). They also released the training environment (35 rebuilt apps, over 110 simulated apps in total), about 11,600 workflows, over 2,000 tasks with automated checks, and reproduction instructions. The 27B model reports 79.9 on AndroidWorld (up from 67.1), 50.4 on MobileWorld (up from 35.9) and 59.6 on SPA-Bench real devices (up from 31.9), reportedly behind only Gemini 3.1 Pro on that last test. In a controlled comparison, adding more training apps lifted success on unseen apps from 60.7% to 67.5%.

Researchers and developers get a usable, fully open base for building phone-operating agents, including the environment and data, which is rarer than weights alone. The app-diversity result is the most useful finding: if it holds, teams should spend effort on varied apps rather than simply more data. Impact is mostly in research for now, and nobody has shown yet that this works reliably on the messy range of real apps and devices that a product would face.

The headline "proves ... can rival proprietary models" overstates it: the models still trail Gemini 3.1 Pro, and the scores are the authors' own, not independently checked. The coverage also underplays that the training examples were generated with Gemini 3.1 Pro as teacher, so the open model leans on a proprietary one. No primary links were extracted, so the numbers rest on a news report of the paper.

1. 01Independent reproduction of the SPA-Bench real-device score (59.6) and AndroidWorld score (79.9) by a group outside the author team, using the released weights and instructions, within the next one to three months.
2. 02Whether the app-diversity result (60.7% to 67.5% success on unseen apps) holds when other labs train on the released environment, or when a different teacher model is used instead of Gemini 3.1 Pro.
3. 03Uptake of the release: downloads and forks of the 9B and 27B weights, and whether any product team (for example Oppo, named in the extraction) or rival open project reports building on it or beating its scores.