CUA-S1 model releases

CUA-S1-FORMS is Cua's first open-sourced "System One" (CUA-S1) specialist, released 2026-09-18 under the MIT license (weights, training code and dataset all published). It is NOT an autoregressive text LLM and does not generate field values: it is a tiny (706,048-parameter, ~2.8 MB checkpoint) one-pass option scorer that acts as the decision layer behind Cua Driver for GUI form filling. Given a UI-element context and a set of typed options (document entities plus CHECK / CLICK / SKIP), it returns one probability per option in a single forward pass and scores FILL/CHECK/CLICK/SKIP independently per element. Architecture (per the model card): byte-level embedding plus a 2-layer Transformer encoder (width 128, 4 heads) run over the context and separately over each option, with an attention head where each option queries context tokens and a shared dot product yields one logit per option, then a softmax over the live option count. Cua reports it scored 99.7% vs 83.6% for TypeSafe's hosted jev-latest API on Cua's own form-filling task (self-reported; a known limitation is over-confident SKIP outside its ~55-concept label catalogue). Directly comparable to (and described as "jev-like") the already-tracked Jev by TypeSafe AI. No tracked provider token pricing yet.