CUA‑S1 Small Decision Model for Form Interaction Released as Open Source

A startup called Cua has announced a new narrow‑scope decision model, CUA‑S1, aimed at automating computer‑use tasks that require only a limited set of local choices rather than full‑scale language generation. The model, inspired by Daniel Kahneman’s System 1/2 framework and the Typesafe “Jev” system, takes a snapshot of the current interface and a list of possible actions, then returns a probability for each option. Unlike conventional large language models that generate text token by token, CUA‑S1 scores pre‑defined actions—such as CHECK, CLICK, SKIP, or use a supplied value—allowing developers to trust the model’s output and drive application behavior directly.

Cua’s first release, CUA‑S1‑FORMS, was trained on synthetic data in under 30 minutes and contains 706 k parameters, with a 2.8 MB checkpoint. It predicts whether to use a given value, check a box, click a control, or skip an element for structured form fields extracted from documents. In a head‑to‑head evaluation against the hosted Jev model, CUA‑S1 achieved 99.7 % accuracy on all decisions, 100 % on action steps, and 100 % on steps that left pre‑filled fields untouched, compared to 83.6 %, 96 %, and 74 % respectively for Jev. Local scoring took 7–9 ms, versus 260–280 ms for a hosted Jev call that includes network latency. The team released the synthetic data, training code, and driver integration under an MIT license, inviting feedback from developers building computer‑use agents.

The Cua team positions CUA‑S1 as a middle ground between brittle scripted automation and general‑purpose LLM agents. By delegating narrow, well‑scoped decisions to a lightweight specialist, agents can remain flexible enough to handle novel situations while avoiding the overhead of calling a full‑scale model for every choice. The current focus is on form interactions, but the authors are exploring broader applications where a general agent could offload routine decisions to models like CUA‑S1.