Cua maps Jev into computer use — jev-use and CUA-S1-FORMS
On 19 September 2026, @trycua published the X article Jev, System One models, and the future of computer use (Dillon Dupont & Francesco). It is one of the clearest category explainers since launch: how much of a computer-use loop still needs a general-purpose LLM once the app already knows the allowed actions.
Their loop: the application observes state and builds candidate actions with IDs; a decision model scores them; the app validates, executes via Cua Driver, then re-observes. A high score is not proof the click worked — calibration is about many examples, not this one action.
Two concrete releases:
| Piece | What it is |
|---|---|
| jev-use | Public-preview recipe wiring hosted TypeSafe Jev to Cua Driver. Jev picks an ID; it cannot invent tools, coordinates, or arguments. Includes a deterministic mock path and a live BYOK path. |
| CUA-S1-FORMS | First model in Cua’s CUA-S1 research family: ~706k trainable parameters (~2.8 MB checkpoint), MIT weights/data/code. Scores form elements (use supplied value / check / click / skip) in batch. Built with ideas from jevlike — not a fine-tune of Jev, and not TypeSafe’s calibration stack. |
They also sketch (but do not ship) a perception path where something like OmniParser turns a screenshot into text regions before a text-only decision model scores candidates.
UnverifiedClaim / hygiene: Dillon’s early form-task numbers (specialist ~99.7% overall decision accuracy vs hosted Jev ~83.6%; local scoring ~7–9 ms vs hosted Jev ~260–280 ms round trip) are narrow and measure different boundaries — treat as vendor experiment, not a ModelSystem.One benchmark. The article itself warns against reading Kahneman as a model taxonomy.
Primary source: x.com/trycua/status/2101437979180904640. Code under trycua/cua (~24k stars on the parent repo at check time).