Opposable runs 2B GUI-Owl 1.5 model on Android for on-device app control

Details

Opposable is an Android app that uses a 2-billion-parameter GUI-Owl 1.5 model to read screens and perform taps/types locally without a server connection.
On a Mac M2 CPU, one step takes 7.3 seconds after optimizations, maintaining 27/36 accuracy on a test set, while the vision encoder runs in 236ms on a Galaxy S25 NPU.
The model is quantized to 4 bits (1.05 GB language model, 445 MB vision encoder) and uses early exit when the model's intent matches a single accessibility element, reducing decode time to 0.8–1.3s.
Accuracy on eight held-out tasks improved from 3–4/8 to 6/8 when the model was given one subgoal at a time with verification against the accessibility tree.
The NPU build is currently limited to Snapdragon chips; Pixel, MediaTek, and Exynos devices fall back to GPU or CPU, and the model is weaker than cloud models on unseen apps.
The app is free at getopposable.com, with the 1.5 GB screen model available as a download in settings.

Key point