CUA-S1-FORMS: A Small Open Specialist for Form Filling

CUA-S1-FORMS is the first release in what Cua calls CUA-S1, a family of small, specialized "System One" models for computer use. In the repository README, System 1 is an engineering analogy for fast, bounded decisions, such as which value belongs in a field or whether to leave an element alone, and Cua says it does not replace a general-purpose agent's planning and reasoning. The forms model scores decisions from structured interface elements and document values instead of generating a response token by token. Application code puts the actions in order, and the optional Cua Driver integration carries them out within explicit action boundaries.

That division of labor is what Cua is testing. In a typical agent, a general-purpose model would generate each next tool call while it works through a form. Cua's longer write-up explains that this model only judges each element against choices the application defines: use a supplied value, check, click or skip. It cannot invent values, read screenshots or extract values from the source document itself, and Cua calls the current release an English-centric research prototype that should not be expected to work reliably on arbitrary forms. The same write-up reports an early head-to-head against TypeSafe's Jev decision model, with Cua's caveat that it was a narrow decision-level experiment.

Cua's announcement presents the release as open-sourcing. The README describes the GitHub component as an early, source-only research release with MIT-licensed model code, synthetic-data generation, training and evaluation, while model weights are hosted separately on Hugging Face, and it tells users to check each model and dataset card for scope, limitations and license.