CUA-S1-FORMS uses 700,000 parameters for form decisions, releases weights and training pipeline

The Cua team introduced CUA-S1-FORMS on Hacker News on September 19, assigning field matching in form interactions to a small model with approximately 706,000 parameters and a raw checkpoint of about 2.8 MB. It reads structured interface elements and values extracted from documents, then chooses whether to fill, check, click, or skip. It does not process screenshots or generate field contents itself. The team has released the training and execution integration code. [Team explanation](https://news.ycombinator.com/item?id=49767564)

The model uses byte embeddings and a two-layer Transformer encoder with a width of 128 and four attention heads, followed by an option-attention mechanism that scores candidate actions. Form elements can be scored in batches, but each element is evaluated independently; the application determines the order of filling fields, checking boxes, and submitting. This division of responsibilities lets engineers delegate repetitive decisions to a local classifier while retaining control over workflow and execution. [Model card and weights](https://huggingface.co/cua-ai/cua-s1-forms)

Training uses 10,000 synthetic scenarios incorporating synonymous labels, similar fields, and distractor data, with training and test sets partitioned by complete field combinations. The public materials also include 196 decisions drawn from three demonstration form pages and three PDFs; the model card reports that the model answered all of them correctly in this small example-based evaluation. The denominator here is individual decisions, so the result cannot be treated directly as an end-to-end form completion success rate across websites and languages. [Dataset](https://huggingface.co/datasets/cua-ai/cua-s1-forms), [evaluation results](https://huggingface.co/cua-ai/cua-s1-forms)

Replication tests should separately track incorrect field selection, skipped required fields, duplicate submissions, and whole-form completion rates, so that large numbers of easy skip decisions do not obscure the few errors that change data. The team measured local form scoring at approximately 7–9 milliseconds, but this excludes the time required for the full document parsing, interface observation, and actual interaction process. The cloud-based Jev used for comparison was not fine-tuned for the same task, and its call times include network latency, so these results cannot establish a speed or capability ranking among general-purpose models. [Test conditions](https://news.ycombinator.com/item?id=49767564)

Deployment still has execution constraints: by default, the source code only generates a plan. Execution and submission must be enabled separately, and the window must be observed again after each modification. Fill operations are rejected if the Driver does not declare compatible field-writing capabilities. [Execution interface documentation](https://raw.githubusercontent.com/trycua/cua/main/libs/cua-s1/README.md)

Chinese-language users particularly need to retest the model: its labels are predominantly in English, and context and options are truncated at 224 and 96 bytes, respectively. These limits leave room for fewer multibyte Chinese characters. Areas to watch include tests involving Traditional Chinese fields, unseen layouts, and error recovery, as well as whether the scores can reliably support abstention thresholds. [Model limitations](https://huggingface.co/cua-ai/cua-s1-forms)