Cua · 2026-09-18 · notable
CUA-S1 — Cua opens the training code for tiny form-filling models
CUA-S1 is Cua's new family of small, specialist computer-use models. The first profile, cua-s1-form-v0, picks form actions without generating text. Cua shipped the code under MIT but is holding back the weights.
A source-only research project for small computer-use models that choose a bounded form action instead of writing JSON.
What is it?
CUA-S1 is a new component inside Cua's computer-use repository holding what the team calls a family of small, specialist computer-use research models. The first profile, cua-s1-form-v0, is scoped to form-oriented interface tasks: given document entities and accessibility data, it chooses which bounded form action to take. Cua frames it as the opposite of a general computer-use agent.
How does it work?
The cua-s1-form-v0 profile is a byte-level transformer encoder with an option-attention classification head, so it scores the permitted options rather than generating tokens. It works from snapshot-bound element tokens and exact targets supplied by Cua Driver, and it separates planning from execution — planning is the default and submitting a form needs explicit authorization. Checkpoints load through safetensors with separate JSON config and checksums.
Why does it matter?
Small decision models are having a moment, and Cua is the first to attach one to a widely used computer-use stack rather than a standalone demo. Notably, Cua published no weights, no datasets and no accuracy claims — the RFC says hosted-provider comparisons and checkpoint accuracy stay out of scope until the evidence is independently reproducible. You get the synthetic-data generator, training and evaluation code to reproduce it yourself.
Who is it for?
agent researchers and computer-use engineers
Try it
uv sync --project libs/cua-s1/python --extra pdf --group test