Cua's small System One models

It proposes; the app executes. The specialist emits only FILL, CHECK, CLICK or SKIP, with values taken from a supplied document, and the application validates and orders the plan: fields first, then checkboxes, then submit only when authorized. Cua Driver carries out the actions in that order.

Reasonably open for a specialist: MIT-licensed source covering synthetic-data generation, training, evaluation and Cua Driver integration, plus weights and a dataset on Hugging Face. What is missing is evidence - no accuracy numbers, no error rate on malformed documents, and no benchmark against the LLM agent beyond side-by-side videos.

The parameter count comes from a third-party post about CUA-S1-FORMS; the announcement itself does not state a size. On beating the agent, all you get is two demo videos shown side by side - no turn counts, no field counts, no success rate on forms the model has not seen.

Only inside a bounded frame. It can emit just those four known verbs, it needs field values handed to it as a document, and the post is explicit that the application validates the plan and that submit happens only when authorized. Nothing is documented about recovering from an unrecognized control or a stale page.

The first in a family. Cua introduces CUA-S1 as a set of small, specialized System One models built for computer use, and calls this form-filling specialist the first of them. The closing post asks companies with repetitive, bounded, high-volume computer workflows what CUA-S1 should learn next, so more specialists are implied.

From a document you supply, not from the model. The specialist only chooses among FILL, CHECK, CLICK and SKIP, the field values are read from the supplied document, and the application then validates and orders the plan before anything is submitted. Generation and decision are kept apart.

A specialist against a general agent on the same form. On one side a small specialist decides how to fill the whole form in a single pass before Cua Driver carries out the actions in order; on the other an LLM agent works through the same form step by step using Cua Driver. No scores or turn counts accompany the comparison.