# OpenMobile-2 releases open smartphone agent models trained across more than 110 apps

published: 2026-10-25T15:05:37.493Z
topic: browsers
source: https://the-decoder.com/openmobile-2-proves-open-source-smartphone-agents-can-rival-proprietary-models/
url: https://aiagentnews.io/stories/openmobile-2-releases-open-smartphone-agent-models-trained-across-more-than-110

Researchers from Nanjing University, SenseTime and other institutions released OpenMobile-2, two open smartphone agent models built on Qwen3.5 and Qwen3.6, along with their training environment and data. The 27-billion-parameter model posted the highest score among open models on SPA-Bench running on real Android devices.

The team released OpenMobile-2 models with 9 billion and 27 billion parameters, built on Qwen3.5 and Qwen3.6, together with the full training environment, data and instructions. Because training on real apps is costly and hard to repeat, the researchers rebuilt 35 apps, mostly popular services in China such as Taobao, JD, DiDi and Meituan, and combined them with existing simulations and emulator apps for a total of more than 110. The simulations reset automatically, and the agent sees only screenshots.

The team generated about 11,600 example workflows and more than 2,000 practice tasks with automated success checks, using Google's Gemini 3.1 Pro as the teacher model. The larger model rose from 67.1 to 79.9 points on AndroidWorld and from 35.9 to 50.4 on the longer-task MobileWorld benchmark. The biggest gain came on real devices, where it climbed from 31.9 to 59.6 points on SPA-Bench.

With the amount of data held constant, adding more apps raised success on unseen apps from 60.7 to 67.5 percent. More than 50 simulated apps also expose direct commands such as "add to cart"; agents used these well only after targeted training, which added up to 10.3 percentage points and about five fewer steps. The authors call this approach experimental, and reinforcement learning produced only limited gains.

## Why it matters if you build agents

Developers building phone agents now have open model weights and a reproducible training environment to test and fine-tune against, rather than relying only on closed models. The results suggest that varying the apps in training matters more than adding more examples from a few apps. Teams should note that the training data came from a closed teacher model and that the direct-command gains exist only in simulated apps for now.