Free AI that taps through phone apps for you now rivals paid ones

Researchers have released OpenMobile-2, an open AI system that uses apps on a phone the way a person would, by looking at the screen and tapping. On real Android phones, the larger version scored 59.6 on a test called SPA-Bench, up from 31.9. The team says that is the best result among open models. The models, training data and instructions are all being shared.

What was built

The work comes from Nanjing University, the AI company SenseTime and several other universities. They released two models, one with 9 billion parameters and one with 27 billion. Parameters are a rough measure of a model's size. The models are built on Qwen3.5 and Qwen3.6.

Training an agent on real apps is hard. Logins, real purchases and the inability to reset an app make practice costly and hard to repeat. So the team rebuilt 35 apps, including Taobao, JD, DiDi, Meituan, Slack and Google Drive. Together with other simulations and an Android emulator, the practice ground covers more than 110 apps.

The results

The agent sees only screenshots, with no access to an app's inner data. Here is how the larger model did on the tests:

• AndroidWorld: 79.9 points, up from 67.1. The earlier OpenMobile-1 scored 64.7.
• MobileWorld, with longer tasks across several apps: 50.4, up from 35.9.
• SPA-Bench on real Android devices: 59.6, up from 31.9.

The smaller model also improved on AndroidWorld, from 54.3 to 74.

Variety beat volume

The main lesson is that the number of different apps matters more than the amount of data. The team kept the data amount fixed and only added more apps. Success on apps the agent had never seen rose from 60.7 to 67.5 percent.

A further training step, called reinforcement learning, helped only a little. It worked mostly in settings like the training setup. On AndroidWorld, the larger model scored 79.9 afterward, slightly below the 80.8 it had before.

Caveats

The models are not fully independent of closed AI. Google's Gemini 3.1 Pro, a closed model, served as the teacher that produced the example workflows. The team generated about 11,600 of them, plus more than 2,000 practice tasks with automatic success checks.

One experiment gave the agent direct commands such as "add to cart," so it could skip menus. After targeted training, agents solved up to 10.3 percentage points more tasks and needed about five fewer steps. These shortcuts exist only in the simulated apps, and the researchers call the approach experimental.

The wider field

Others are working on the same problem. Oppo has an open agent called X-OmniClaw that runs on real devices and replays paths it was shown in apps like Taobao or Meituan. Among closed models, Google has built phone control directly into Gemini 3.5 Flash. Phone assistants that handle app tasks are becoming cheaper and more open, so they may soon show up in tools you or your business use.