CUA-S1 on Hacker News

Only partly, and the authors say so. They note the specialist was trained specifically for the form task and its skip convention while hosted Jev was not fine-tuned for it, so they frame it as an experiment in scoped specialization rather than a head-to-head.

No. The post states the local scoring time and the hosted call including network latency measure different things and are not end-to-end form completion times. One is a local forward pass, the other an API round trip.

Not on its own. The release covers forms only, driven by structured elements and values pulled from a document; it does not predict new values for text fields and does not look at screenshots. The first training run took under 30 minutes on synthetic data.

A router over many narrow specialists. The central comment asks whether a parent model would pick among specialists for forms, Wikipedia or video editing, possibly with a model choosing the models, and another argues for one capable package instead of disaggregation.

706k parameters, with the original checkpoint at 2.8 MB and the first training iteration under 30 minutes on synthetic data. Given structured elements and values pulled from a document, it predicts whether to use a value, CHECK, CLICK, or SKIP per element. Element decisions are scored together and your code orders them, with Cua Driver executing one at a time.

The central request is a router over many narrow specialists, with a parent model choosing among them for forms, Wikipedia or video editing, possibly with another model choosing the models. One reply argues for the opposite, that capable-but-efficient should live in one package rather than being disaggregated, pointing at the divide between viewing a scene and acting in it.

Yes, with the synthetic data generation, training, evaluation, and Driver integration released under libs/cua-s1 on an MIT licence. Built from ideas and code in jevlike, which is a useful pointer if you want the scoring approach without the computer-use framing.