Cua trains a 2.8 MB scorer to replace the per-field reasoning step in form agents

706,048 parameters at four bytes each comes to 2,824,192 bytes, about 2.8 MB ^[1]. The published checkpoint is 2.8 MB ^[3], so the file holds float32 weights and very little else.

The model returns probabilities, not actions ^[8]. Cua Driver selects from them and orders the work: fields first, then checkboxes, and a submit button only when authorized ^[9]; that authorization check is a line of driver code.

The saving is confined to the step the model takes over. CUA-S1-FORMS can select only values an upstream extractor has already supplied as labeled entities, and it cannot infer or compose a value that is missing ^[18]. Whatever reads the referral letter still runs. What leaves the loop is the inspect-reason-act cycle a general agent repeats for every element ^[10]. Cua's release measures accuracy; per-decision latency and cost go unreported, for the model as for the extractor ^[5].

For the real-form result to mean anything on someone else's forms, a few things have to hold. The 196 decisions came from three forms and three PDFs, roughly 65 decisions per form ^[15][3], and a single wrong call would have put the score at 99.5% ^[3]. The training generator drew from 55 field concepts ^[13], the label vocabulary is English-centered ^[17], and the model card says the model has not been validated beyond the demonstration set ^[17]. Testing covers those 55 concepts, and extractors that spell their entity labels the way the generator does.

The generator deliberately plants confusing pairs, such as a primary phone number next to an emergency contact number ^[13].

Cua describes the model as "jev-like" ^[19] and reports 99.7% against 83.6% for TypeSafe AI's hosted jev-latest API on Cua's own task ^[21], a gap of 16.1 points ^[4]. Cua also states that it trained its model to recognize already-completed fields as no-ops and that the Jev service it tested was not trained for that behavior ^[21]. On a blank form the distinction does not arise. TypeSafe pitches Jev as a general decision model for classification, routing and other software workflows, while Cua trained one specialist and released the weights, the data generator, the evaluation code and the driver integration ^[20].

The people doing this have measured agents before. Bonacci and Cua researcher Dillon DuPont were among the authors of Microsoft's Windows Agent Arena, a benchmark for evaluating agents inside a real Windows environment ^[5]. Bonacci founded Cua in 2025 after leaving Microsoft, starting with Lume for running macOS virtual machines on Apple Silicon, and the company joined Y Combinator's Spring 2025 batch ^[6].

The published dataset is 234,067 rows ^[14]. The named splits, roughly 150,000 training, 18,000 validation and 20,000 test examples, come to about 188,000, some 46,000 rows short of the total ^[14][2].

Because the weights, generator and evaluation code are MIT ^[4], the 196-decision harness can be re-pointed at in-house forms in an afternoon.