OpenMobile-2 Brings Open-Source Smartphone Agents Closer to Proprietary Rivals
Researchers from Nanjing University, SenseTime and partner institutions have released OpenMobile-2, comprising 9-billion-parameter and 27-billion-parameter smartphone agents based on Qwen3.5 and Qwen3.6, respectively. The package includes training data, instructions and a reproducible Android environment covering more than 110 applications. It features 35 carefully rebuilt versions of services such as Taobao, JD, DiDi, Meituan, Slack and Google Drive, with persistent cross-app workflows and automatic resets.
The environment produced about 11,600 example workflows and more than 2,000 automatically verified practice tasks, using Gemini 3.1 Pro as the teacher model—meaning the open agents still learned from a proprietary system. On AndroidWorld, the larger model rose from 67.1 to 79.9 points and the smaller from 54.3 to 74, while OpenMobile-1 scored 64.7. The larger model also improved from 35.9 to 50.4 on the multi-app MobileWorld benchmark and from 31.9 to 59.6 on real-device SPA-Bench, which the researchers say is the strongest result among open models.
Controlled tests suggest breadth matters more than volume: with the same data budget, expanding the number of training apps raised success on unseen apps from 60.7% to 67.5%. Reinforcement learning delivered limited gains and slightly reduced the larger model’s AndroidWorld score from 80.8 to 79.9. Experimental direct commands in simulated apps improved completion by up to 10.3 percentage points and shortened tasks by about five steps after targeted training.
Continue in the app — vote & join in ➔