CUA-S1-FORMS

CUA-S1-FORMS is the first release in what Cua, a San Francisco company, calls its CUA-S1 family of System One models: small, specialised models for the bounded decisions an agent makes constantly while operating a computer. It was published on 18 September 2026 with 706,048 trainable parameters and a 2.8 MB checkpoint, under the MIT licence, with the training code and the synthetic dataset released alongside the weights. It does not generate text. Given the context of a form and a set of extracted document values, it scores every permitted option in a single forward pass and returns a probability for each, which means the decision stays inside ordinary code rather than becoming another call to a large model. The architecture is byte-level: a two-layer transformer encoder of width 128 with four attention heads, taking 224 bytes of form context and 96 bytes per option. It was trained on 10,000 synthetic episodes of forms with two to sixteen fields, split so that no form signature appears in both training and test. Cua reports 99.95% top-1 accuracy on that synthetic test, 100% across 196 decisions on three real forms and three PDFs, and 37% on a shuffled-context control that exists to prove the model is reading the elements rather than guessing. It also publishes a comparison — 99.7% against 83.6% for a hosted general model on the same task. None of those figures is stored as a benchmark score on this board, for two reasons the company and its coverage both state: the evaluation is Cua's own and the real-form set is small and tailored, and the dated independent write-up of the release notes that the comparison favours the specialist, because Cua trained its model to recognise already-completed fields as no-ops and the model it was compared against was not. The honest limits are on the model card: it can only choose among values already supplied as labelled entities, so it cannot infer or compose anything missing; it was trained on synthetic data with little real-world validation; and the encoder is English-focused. Because it runs locally there is no per-token price at all.