1 · The screen
2 · The decision
| Option | Taken | p |
|---|
cua-s1-4b-0.2 is a LoRA on Qwen3.5-4B. It is given a goal, one screen and a closed list of (element, action) options, one letter each, and picks the single best next action in one forward pass, on WebGPU. The screens below are real Word, Excel and PowerPoint steps from GUI-360's test split; ✓ marks the action that was taken there, and the rows the model would act on are in bold. Cua trained two adapters: one reads the accessibility tree, the other a screenshot through Qwen3.5's vision tower. Each is its own download of about 5 GB.
Accessibility tree · Screenshot
Loading the model…
| Option | Taken | p |
|---|