CUA-S1: A 706k-parameter specialist that scores form actions instead of generating tokens

Show HN: CUA-S1 – A System One Model for Computer Use

Cua open-sourced CUA-S1-FORMS, a 706k-parameter model (2.8 MB checkpoint) that scores form-element actions—FILL, CHECK, CLICK, or SKIP—instead of generating tokens. Trained on synthetic data in under 30 minutes, it hit 99.7% accuracy on their form decision set vs. 83.6% for hosted Jev. Local scoring takes 7–9 ms per form, compared to 260–280 ms per remote call. It does not handle screenshots or predict new text values. The team frames this as a specialist for decisions that are too variable to script but too narrow to warrant a general-purpose LLM call.