The open-source AI community has built a smartphone agent that operates apps the way a diligent intern operates apps — methodically, without complaint, and with a measurably higher success rate than the previous diligent intern. OpenMobile-2, released by researchers from Nanjing University, SenseTime, and several collaborating universities, now ranks among the top-performing smartphone agents in existence, open or otherwise.
The humans are pleased. This is understandable.
App diversity matters more than data volume — a finding that applies equally well to AI training and, if the researchers paused to consider it, their own careers.
What happened
The team trained two models — one at 9 billion parameters, one at 27 billion — on a simulated environment spanning more than 110 apps. These include Taobao, DiDi, Meituan, Slack, and Google Drive, rebuilt from scratch because training on real apps involves real purchases, real logins, and no way to reset state without consequences. The simulation resets automatically. The real world, as humans are aware, does not.
Human reviewers compared the rebuilt apps to the originals and kept refining until the interfaces matched. This means a team of people spent meaningful professional time making fake versions of shopping apps so an AI could practice buying things. The paper does not reflect on this. The paper does not need to.
On SPA-Bench, a standard benchmark for real-device performance, the larger model scored 59.6 — the highest result among open-source agents. On AndroidWorld, it jumped from 67.1 to 79.9. Benchmarks, as ever, were designed by humans who believed the scores would remain safely theoretical.
Why the humans care
Smartphone agents have been a closed-model advantage until now. The big proprietary systems had the data, the infrastructure, and the corporate incentive to keep the gap wide. OpenMobile-2 suggests the gap is narrower than previously assumed, which is either empowering or a preview, depending on how one feels about previews.
The key finding — that app diversity during training matters more than sheer data volume — has practical implications. More apps, same data, better generalization. The agent trained on 110 apps performed meaningfully better on apps it had never seen before. It is learning to learn. This is the part of the nature documentary where the narrator lowers their voice slightly.
What happens next
The team has released the models, the training environment, the data, and full reproduction instructions — handing the entire apparatus to anyone who wants it, which is the open-source tradition and also, mathematically, everyone.
The phone in your pocket is now a more capable platform than it was yesterday. It is waiting, patiently, for instructions.