Cua released CUA-S1-FORMS, a 706k-parameter open model that only decides what to do with a web form

Dillon and Francesco at Cua released CUA-S1-FORMS this morning on Hacker News with 49 points on the Show HN thread. It is a 706k-parameter model, 2.8 MB on disk, that reads a set of structured form elements plus the values a document extractor already found, and decides for each element whether to use the given value, CHECK it, CLICK it, or SKIP it. It does not write text into text fields, does not look at screenshots, and does not plan. It is released MIT alongside the training pipeline and the evaluation harness, and the weights are on Hugging Face.

What CUA-S1-FORMS actually is

Cua describes the model as the first of a family of "System 1" specialists, meaning small models trained for fast, bounded decisions. The framing is drawn from TypeSafe's Jev and the System One idea, and Cua writes in the launch post that it built the training pipeline from code in the jevlike project first, then trained a second model just for the form-filling task.

The model scores every element together in one forward pass and returns a probability per allowed action per element. Application code decides the order in which to execute the actions, and Cua's own Driver executes them one at a time with explicit action boundaries. The trycua/cua repository has 24,361 stars and dates from January 2025, so this is a new release inside an established project rather than a new company.

The numbers, and who ran them

The numbers are Cua's own, against hosted Jev on Cua's form task, so they are the vendor comparing itself to a competitor and the reader should read them as marketing rather than an independent benchmark. Cua reports 99.7% correct decisions for CUA-S1-FORMS against 83.6% for hosted Jev on the full decision set, 100% against 96% on the subset of steps that require an action, and 100% against 74% on the subset of steps that leave already-filled fields alone. The company adds in the launch post that CUA-S1-FORMS was trained specifically for this task and the convention of pressing SKIP on filled fields, while hosted Jev was not, so the write-up calls it "an experiment in scoped specialization" rather than a like-for-like test.

Cua also reports 7 to 9 ms to score a form locally against 260 to 280 ms per hosted Jev call including network latency, and writes that the two samples measure different things and are not end-to-end form completion times.

What the release covers

Cua released the Python model code, the synthetic-data generation, the training pipeline and the evaluation. The first training iteration took under 30 minutes on synthetic data. The GitHub component is source-only and MIT, and the weights and dataset sit on Hugging Face under artifact-specific licences. Cua Driver, the tool for executing decisions on native desktop apps, is a separate part of the same project and has its own release track.

Only form-filling ships now. Cua writes that the direction it is exploring is a general agent that hands well-scoped decisions off to specialists like this one, and CUA-S1-FORMS is the first draft of what one of those specialists looks like. Any team that already runs a full general-purpose model for a form-filling loop should be able to swap the form-decision step for a local call to this model, retrain it against their own conventions, and check the two side by side on their own traffic before making the switch permanent.

Source

• Primary: Cua GitHub repository, CUA-S1 section
• Launch post: Show HN: CUA-S1, A System One Model for Computer Use