CUA-S1-FORMS

CUA-S1-FORMS is a small specialised model from Cua, a San Francisco company, published on 18 September 2026 as the first release in its CUA-S1 family. It does not generate text. Fed the context of a form plus document values already pulled out, it scores each permitted option in one forward pass, so the decision never needs a second call to a large model.

Features

• Single-pass choice: returns a probability for each permitted option in one forward pass.
• Byte-level encoder: a small two-layer transformer reads the form context and each option as bytes.
• Synthetic training: trained on synthetic form episodes with two to sixteen fields, with no form appearing in both training and test sets.
• Local operation: runs on the machine doing the work, with no API call.
• Open release: the weights, training code and synthetic dataset are published under the MIT licence.

Clients

Developers building agents that operate computers and fill in forms from extracted document values.

Key facts

• Parameters: 706,048 trainable parameters, with a checkpoint of 2.8 MB.
• Limits: it can only choose among values already supplied as labelled entities, and it was trained on synthetic data with little real-world validation.
• Language: the encoder is English-focused.