After reading this, the reader knows how CUA-S1-Forms divides model judgment from interface execution and where its evaluation is still too narrow.

Cua Releases a Specialist Form Model

The 706,048-parameter checkpoint selects document values and form actions in one pass, while ordinary code controls ordering, execution and verification.

Cua has released CUA-S1-Forms, a small model designed for one narrow part of computer use: deciding what value or action belongs in each element of a form.

The model does not look at a screen and freely plan a task. It receives structured text describing one interface element and a list of permitted options. Those options can be document entities, such as a phone number, or fixed actions such as check, click and skip. It returns one probability per option.

Application code remains responsible for ordering the selected actions and sending them to Cua Driver. Fills happen before checkboxes and the submit click. That separation is deliberate: the model makes bounded choices, while deterministic software controls the workflow.

Why it matters

Computer-use agents often ask one large model to perceive an interface, plan a sequence and execute every action. When something fails, it can be hard to distinguish a perception error from a planning error or an execution bug.

CUA-S1 proposes a different architecture. A tiny specialist handles repeated local decisions, and a larger agent or conventional program handles the broader task. The approach may reduce latency and make some failures easier to inspect because each element has an explicit list of choices.

The checkpoint is unusually small by current standards. Its model card describes a byte-level encoder with two Transformer layers, width 128 and four attention heads. It has 706,048 trainable parameters and a 2.8 MB checkpoint. It scores every actionable element independently and can batch the elements in parallel.

Training used 10,000 synthetic episodes built from forms with two to 16 fields, a catalogue of 55 concepts, distractor entities and deliberately similar fields. Test splits were separated by exact form signature so an identical field set would not appear in training and testing.

Strong numbers, narrow evidence

Cua reports 99.95% top-one accuracy on roughly 15,000 synthetic test decisions and 100% on a real demo evaluation. The real set contains only three forms, three PDFs and 196 decisions. That is useful evidence that the pipeline can work, but it is far too small to establish performance across arbitrary websites, languages and document formats.

The project also reports 99.7% against 83.6% for the hosted Jev API on the same task. The comparison is not a general claim that CUA-S1 is a better model. CUA-S1 was trained for this form convention, including recognizing an already-filled field as a no-op, while Jev was used without task-specific fine-tuning.

The limitations are substantial. The model only chooses among values already extracted as Label: value pairs. It cannot invent a missing value, and its vocabulary is English-centric. Unfamiliar layouts, ambiguous labels or a bad document parser can still produce the wrong action.

Cua’s security guidance recommends isolated environments, least-privilege credentials, action limits and independent outcome checks. Consequential or irreversible operations should require human confirmation. These controls matter because a high benchmark score does not turn a form-filling model into an authorization system.

The release is best understood as an architectural experiment made reproducible: a small, inspectable decision layer placed between extracted data and guarded interface actions. The next evidence should come from larger, independent real-world evaluations with varied layouts and failure reporting.

Verification

  • Tier 0 — VERIFIED: CUA-S1-Forms weights, card and code are public under MIT terms. Primary sources: https://huggingface.co/cua-ai/cua-s1-forms and https://github.com/trycua/cua
  • Tier 0 — VERIFIED: The model card specifies 706,048 parameters, a 2.8 MB checkpoint and a one-pass option-scoring interface. Primary source: https://huggingface.co/cua-ai/cua-s1-forms
  • Tier 1 — DEVELOPER-REPORTED: Synthetic, real-demo and Jev comparison results come from Cua and were not independently reproduced here. Same primary source.
  • Tier 0 — VERIFIED: Cua documents the small real evaluation, English-centric vocabulary and restricted extraction format. Same primary source.
  • Tier 2 — ANALYSIS: Architectural and deployment implications are editorial analysis.

Glossary candidates

  • Option scorer: A model that assigns a probability to each allowed choice.
  • Fail-closed: Refusing to act when required state or safety information is missing.

Cold-reader sentence: CUA-S1-Forms is a tiny bounded decision model for form filling whose promising results come from a deliberately narrow evaluation.