OpenMobile-2 boosts smartphone agent success rates

Researchers from Nanjing University, SenseTime, and other institutions released OpenMobile-2, a project designed to close the gap between open and proprietary smartphone agents. The team released two agent models built on Qwen3.5 with 9 billion parameters and Qwen3.6 with 27 billion parameters. These models were trained in an environment containing more than 110 simulated apps.

The training environment included 35 rebuilt apps popular in China, such as Taobao, JD, DiDi, and Meituan. Human reviewers compared these rebuilt apps to the originals to ensure workflows and interfaces matched. The team generated about 11,600 example workflows and more than 2,000 practice tasks with automated success checks. Google's Gemini 3.1 Pro served as the teacher model that produced the examples for this training process.

Increasing the number of apps during training pushed the success rate on unseen applications from 60.7 to 67.5 percent when data volume was constant. On the MobileWorld benchmark, the larger model improved from 35.9 to 50.4 points. The predecessor OpenMobile-1 scored 64.7 on the AndroidWorld test. More than 50 simulated apps offer direct commands like 'add to cart,' allowing agents to skip navigating through menus. Once trained on these direct commands, agents solved up to 10.3 percentage points more tasks while needing about five fewer steps.

A custom benchmark built by the team contained 215 tasks, nearly half of which span multiple apps. Task completion in this benchmark is judged by a program comparing app data rather than a language model. Microsoft Research classified API agents as more reliable and GUI agents as more flexible in a 2025 study.