A model with one job: filling in forms on a computer screen, field by field. Think of it as the data-entry clerk of the agent world. An AI agent can find a flight, read the reviews, pick the hotel. Then it hits the booking form: name, dates, passport number, country dropdown, 40 fields of pure tedium. That is the step where agents choke today. This model is built for exactly that step: it looks at each field and answers fill, check, click, or skip, at machine speed.
Why a special model for this? A chat model fills a form the way a novelist fills out a customs declaration: it writes an essay, then hopes someone extracts the right boxes from it. This model never writes anything. It is the difference between proofreading a paragraph and ticking boxes on a clipboard: faster, cheaper, and with a confidence number on every tick.
It is tiny: hundreds of thousands of parameters, not billions, so it runs without a GPU the way a pocket calculator runs without a data center.
Status, honestly: this is a research release. No weights are published yet, which means nobody can run it. The 99.7 percent form accuracy you may have seen comes from the vendor launch thread, and the model card itself claims no result. We are tracking it because the job it targets, form filling, is the most boring and most universal bottleneck in computer use. When weights ship, this page gets the real numbers and the run-it-yourself details.