Cua's CUA-S1 scores form decisions with a 706K-parameter model, not an LLM call

Cua, a startup working on computer-use agents, has released CUA-S1, a family of small models built to handle narrow decisions inside agent workflows rather than route every step through a general-purpose LLM. The first model, CUA-S1-FORMS, weighs 706,000 parameters and ships as a 2.8 MB checkpoint. Founders Dillon and Francesco posted the work to Hacker News, where it drew 60 points and a handful of comments from agent builders.

The design borrows from Typesafe's Jev and its System One framing, a reference to Daniel Kahneman's split between fast intuitive thinking and slow analytical reasoning. Instead of emitting tokens one at a time, CUA-S1 takes the current context plus a set of candidate choices and returns a probability for each. The output is a ranked set of options an application can inspect and act on directly, rather than free-form text a developer has to parse and trust.

CUA-S1-FORMS handles one job: given structured elements and values pulled from a document, it predicts whether each element should take the supplied value, be checked, be clicked, or be skipped. It does not invent text for empty fields and does not look at screenshots. Decisions are scored as a group, and the developer's code orders the resulting actions, which Cua Driver then runs one at a time. Training on synthetic data took under 30 minutes for the first iteration.

On Cua's own form benchmark, the specialist beat hosted Jev across the board: 99.7% correct on the full decision set versus 83.6%, 100% on steps requiring an action versus 96%, and 100% on leaving already-filled fields untouched versus 74%. The company is explicit that the comparison is not apples to apples, since the specialist was trained on this exact task and convention while hosted Jev was not fine-tuned for it. Latency numbers point the same direction: 7 to 9 milliseconds to score a form locally against 260 to 280 milliseconds per hosted Jev call including network round trip, though Cua notes these measure different things and are not end-to-end completion times.

The pitch is not that small models replace big ones. Cua frames the target as the space between a brittle script and a full agent loop, where form layouts and field contents shift enough to break hardcoded rules but the decision set stays narrow. The team wants to test a hybrid: a general agent hits something novel, then hands well-understood decisions to a specialist like this one. That direction is under exploration, not shipped.

The release covers forms only, and Cua has open sourced the synthetic data generation, training, evaluation and Driver integration under libs/cua-s1 with an MIT license. The open question is whether the accuracy holds when field conventions differ from the training set, and whether other recurring agent decisions, such as navigation or list selection, can be scoped tightly enough to justify their own specialists.