A 706K-parameter form filler

That objection is in the thread itself, from a reply arguing that 23 turns says more about the agent harness than about the model. So the honest reading is a latency demonstration on one form, not a head-to-head result - and nothing is published about field count, form variety, or what happens on a page the specialist cannot parse.

No. The post reports one form filled in a 50 ms pass against 39.6 seconds and 23 turns for the LLM agent, and stops there. No field-level success rate, no malformed-input behaviour, no cost figure per fill - the entry's own caveat flags it as author-reported from a single demonstration.

Bounded scope. It is CUA-S1-FORMS, a specialist that selects from FILL, CHECK, CLICK or SKIP and pulls values from a supplied document, with the application validating and ordering the plan. Free-form reasoning, novel layouts and anything outside those four verbs sit outside what the announcement claims.

You can get the weights and dataset from Hugging Face and the source is MIT, but 50 ms is the author's own measurement and the post documents no hardware, no form complexity and no timing method. Treat it as an order-of-magnitude claim, and expect your own number to differ once network time is counted.

Arsh, described as a 16-year-old builder, quote-posted Cua's September 18 announcement of CUA-S1-FORMS with the 706k-parameter and 50 ms claims, and the original post lists 52.3K views.

Cua describes it as a family of System One Models that are small, specialized and built for computer use, with CUA-S1-FORMS as the first one open-sourced, at github.com/trycua/cua.

A reply from Nav argues that if an LLM agent took 23 turns it says more about the harness than the model, and that these kinds of models may be good for short and simple tasks while computer use is often long-horizon. That objection is the main caveat in the thread.