Apple Teaches a Small Model to Navigate Screens Like a Big One

Apple, the company that makes the iPhone, published a paper on using cheap AI agents to interact with smartphone interfaces. The agent works with the interface by tapping the screen and moving between different app screens without any predefined script. It turned out that a model with 8 billion parameters, which is fairly small by today's standards, scored almost 7 times better than a model trained on ready-made demonstrations when handling unfamiliar Android apps. When applied to real iOS apps, this agent beat Google's Gemini-3.5-Flash model and matched the much larger Gemini-3.1-Pro model.

Usually, when we think of AI agents that automate interface work, like tapping around a phone or website screen, they rely on large models at every step of the interaction. That makes these interactions fairly expensive. Instead, Apple used its small model to work with an actual app. The model studied the app and learned from its own mistakes, rather than being fed pre-recorded demonstrations. This made it possible to match the performance of much larger models. Unfortunately, for now this work exists only as internal Apple benchmark results. There is no public release of the model that anyone could deploy on their own.