CUA-S1-FORMS makes one-pass choices among predefined actions, then hands the plan to Cua Driver for execution.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [Cua on X](https://x.com/trycua/status/2101014004927729737)
Why it matters #
Cua is testing whether small, task-specific models can remove expensive LLM calls from repetitive agent workflows while keeping actions constrained by ordinary code.
Francesco Bonacci (@francedot), the founder of Cua, released a small open-source model on September 18th that plans how an AI agent should fill out forms without repeatedly calling a large language model.
CUA-S1-FORMS is the first release in what the San Francisco-based Cua calls its CUA-S1 family of specialized "System One" models. The model has 706,048 trainable parameters and a 2.8 MB checkpoint, making it tiny by contemporary AI standards. Cua published the model, training code and dataset under the MIT license.
Bonacci previously worked at Xbox and Microsoft AI. He 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. Bonacci founded Cua in 2025 after leaving Microsoft, starting with Lume, an open-source tool for running macOS virtual machines on Apple Silicon. Cua joined Y Combinator's Spring 2025 batch and has since expanded into drivers, cloud desktops and evaluation tools for computer-use agents.
The new model narrows that work to a single bounded decision: given information extracted from a document and an element on a form, decide whether to fill the field with one of the supplied values, check a box, click a button or skip the element.
CUA-S1-FORMS does not generate text. It scores the permitted options in one forward pass and returns a probability for each. Cua Driver then translates the selected options into interface actions and orders them so fields are filled first, checkboxes follow and a submit button is clicked only when authorized.
That division of labor moves part of the agent loop out of a general-purpose language model. A conventional agent may inspect a field, reason about it, perform an action and repeat the cycle for every element. Cua's specialist instead evaluates the actionable elements independently in a batch, leaving execution and safeguards to ordinary application code.
A narrow model with a narrow test
The architecture uses byte-level embeddings and a two-layer Transformer encoder with a width of 128 and four attention heads. Cua trained it for six epochs on 10,000 synthetic episodes containing forms with two to 16 fields. The synthetic generator draws from 55 field concepts and deliberately inserts confusing pairs such as a primary phone number and an emergency contact number.
The published dataset contains 234,067 rows, including roughly 150,000 synthetic training examples, 18,000 validation examples and 20,000 test examples. Cua also assembled a 196-decision evaluation using three forms and three PDFs: a referral letter, a resume and a police incident report.
Cua reports 99.95% top-1 accuracy on its form-disjoint synthetic test and 100% on the 196-decision real-form evaluation. Those figures remain Cua's own results, and the real evaluation is too small and tailored to establish performance across arbitrary business forms. The model card explicitly says CUA-S1-FORMS has not been validated beyond the demonstration set and uses an English-centered label vocabulary.
The model also depends on an upstream document extractor. It can select only values already supplied as labeled entities, meaning it cannot infer or compose missing information. That constraint limits generality while reducing one common failure mode: inventing a field value that was never in the source document.
Cua follows Jev with an open implementation
Cua's release arrived four days after TypeSafe AI introduced Jev, a managed model that also returns typed decisions and probabilities instead of generating strings. Cua describes CUA-S1-FORMS as "jev-like" and uses the same broad contract of scoring a predefined set of options.
The projects differ in scope. TypeSafe pitches Jev as a general decision model for classification, routing and other software workflows. Cua trained its first specialist around one workflow and released the weights, data generator, evaluation code and Cua Driver integration.
Cua reports that CUA-S1-FORMS scored 99.7% against 83.6% for TypeSafe's hosted jev-latest API on Cua's task. The comparison favors the specialist on at least one important convention: Cua trained its model to recognize already-completed fields as no-ops, while the tested Jev service was not trained for that behavior. Cua also says CUA-S1-FORMS does not reproduce TypeSafe's reinforcement-learning method or its claimed probability calibration.
The release gives Cua a model layer above the desktop infrastructure it has spent the past year building. As of September 19th, Cua's main open-source repository had about 23,300 GitHub stars. Cua is now asking businesses with repetitive, bounded computer workflows to suggest the next specialist for the CUA-S1 family, using open models to pull more workloads toward the driver and cloud infrastructure where Cua already operates.