Jev had just been popular for 5 days when Cua compressed the decision model to 2.8MB.

Beating AI News Flash: Cua has open-sourced CUA-S1-FORMS, which has only 706,000 parameters and is 2.8MB in size, specifically designed for automatic form filling. The system first extracts information such as names and phone numbers from PDFs, then the model determines which fields this information should be filled into, and finally hands it over to Cua Driver for execution.

It also follows the System One route recently popularized by Jev, only making selections and judgments, without generating text. But Jev is aimed at more general decision-making tasks, while CUA-S1-FORMS narrows the scope further, only solving information matching within forms.

In Cua's self-testing, CUA-S1-FORMS scored 99.7% on this set of form-filling tasks, while Jev scored 83.6%. This result has an obvious home-field advantage, as the former was trained specifically for form filling, and it cannot be understood as a 700,000-parameter model comprehensively defeating Jev.

Jev wants to replace part of the large-model decision-making in Agents, and Cua continues to break it down further. When the task is sufficiently fixed, a general decision-making model is not even needed.