Show HN: Computer use on the phone itself: a 2B model that operates Android apps (getopposable.com)

A new Android app, Opposable, utilizes a 2-billion-parameter model called GUI-Owl 1.5 to process and interact with other apps directly on mobile devices without needing a server. This model employs accessibility features to read the screen and execute actions, significantly enhancing mobile computing capabilities. On devices like the Samsung Galaxy S25, the app harnesses a specialized vision encoder that dramatically reduces processing time, completing tasks around 16 times faster than standard desktop CPUs. Specifically, the model operates in 4-bit quantization, optimizing memory use while delivering competitive accuracy on a range of common tasks, such as setting alarms or texting contacts. For the AI/ML community, this development is significant as it represents a significant leap in mobile AI efficiency and usability, enabling real-time interaction with applications through sophisticated models. The GUI-Owl 1.5's architecture allows it to learn and execute complex tasks directly on the device, preserving user privacy by minimizing reliance on external servers. While the model demonstrates impressive performance on supported Snapdragon devices, its current lack of compatibility with other chipsets highlights potential growth areas for cross-platform AI deployment.