Opposable's Android screen model takes 7.3 seconds per step
About 40 common tasks run as fixed code, while the on-device loop takes 7.3 seconds per step on a Mac M2 CPU.
| Opposable | |
|---|---|
| model | GUI-Owl 1.5 2B |
| quantization | 4 bits |
| weights | 1.05 GB + 445 MB |
| step time | 7.3 s on M2 CPU, 4 threads |
| previous step time | 10.6 s on M2 CPU |
| encoder | 236 ms on Galaxy S25 NPU |
| accuracy | 27/36 on 36-screen set |
| pruning accuracy | 28/36 |
The 7.3-second figure comes from a Mac M2 CPU with 4 threads. It measures one screen step, not a finished task.
The vendor's 36-screen test scores 27 correct. That leaves the model as a slow fallback for tasks the fixed-code path cannot cover.
The page does not show iPhone latency, task completion outside the 36-screen set, or the cost of a failed step. It also does not say whether the 1.05 GB language model and 445 MB encoder fit on lower-end phones.
Android uses accessibility services and screenshots to feed GUI-Owl 1.5 2B through llama.cpp. iPhones require a companion computer.
The 7.3-second step and 27/36 score make it a slow fallback for tasks fixed code cannot cover.
After getopposable.com. We did not report this. The pictures, if any, are theirs.
