cua-s1-forms

cua-s1-forms

Scores the possible actions for each element of a GUI form, such as fill, check, click or skip, and returns a probability for each. A tiny model for developers building form-filling agents.

Made by cua-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: cua-s1-forms is a tiny model, with 706,048 parameters, for filling in forms on a screen. It does not generate text. You hand it a form element along with the choices on offer, and it replies with a likelihood for each choice at the same time. It belongs to the Cua-S1 research family from cua-ai.

Who it suits: Developers building agents that fill forms from documents, and researchers who study small models that choose UI actions.

Industry | Who uses it | How they use it | Example
Software development | Developers building computer-use agents | Score every field on a form to choose an action for it | Deciding whether a phone field should be filled, skipped or clicked
Research | Researchers on small action models | Study how a tiny model ranks options for form fields | Testing a 706,048 parameter model on form checks
Back-office automation | Teams that move data from documents into forms | Pair the model with code that fills fields and submits | Filling a registration form from a scanned document

Skip it if you want a model that writes or chats, since it only scores options.

Where to use it: It runs in Python through the cua-s1 package, and the files download on Windows, macOS or Linux.

• Python: where the checkpoint is loaded
• The cua-s1 package: provides the loader and the device helper
• A Cua Driver session: where the planner loop sends its actions

Use case | Who | What they get back
Form field scoring | Agent developers | One probability for each option on a field
Action choice | Automation builders | The highest scored option to execute
Small model research | Researchers | A compact checkpoint to inspect and test

How to set it up: You need Python, the cua-s1 package and the two safetensors files. Load only that safetensors pair. The older pickle checkpoint was removed, and the card warns that unpickling a checkpoint can run arbitrary code.

• Python with pip
• The cua-s1 package, installed from a checkout of the Cua project
• The two files, a safetensors file and a json file

1. Install huggingface_hub.
2. Get the Cua project code and install cua-s1 from it.
3. Download only the two files, leaving out any pickle copy.
4. Load the checkpoint with the cua-s1 loader.

pip install -U huggingface_hub
hf download cua-ai/cua-s1-forms

The card installs the package like this, from a checkout:

uv sync --project libs/cua-s1/python

How to use it: Download the two files and load them with the loader. The loader checks the format, version and signature before it returns the model.

from pathlib import Path
from huggingface_hub import snapshot_download
from cua_s1.model import load_checkpoint, select_device

root = Path(snapshot_download(
    "cua-ai/cua-s1-forms",
    allow_patterns=["cua-s1-forms.safetensors", "cua-s1-forms.json"],
))
model, collator, config = load_checkpoint(root / "cua-s1-forms.safetensors", select_device("auto"))

With the command line, leave pickle files out when you pin an older revision:

hf download cua-ai/cua-s1-forms --exclude "*.pt" --local-dir ./cua-s1-forms

Tool | What it does
snapshot_download | Fetches only the two files the model needs
load_checkpoint | Loads the model after checking its format and signature
select_device | Picks the device to run on
hf download | Downloads from the command line, with pickle files excluded

Example: For a phone number field on a registration form, you pass the form context and the options: fill with the phone number from the document, check, click and skip. The model returns one probability for each option, and the executor takes the highest.

For newcomers and experts: If you are new to this, think of it as a ranker. It sees one form field and a short list of choices, and it says how likely each choice is. Start by downloading the two files and loading them with the Python snippet.

If you know the field, it is a byte level model with a 2 layer Transformer encoder, width 128 and 4 heads. Each option queries the context through an attention head, and a softmax then spreads probability across the options still in play. The card says it was trained on 10,000 synthetic episodes, and code outside the model decides the order of actions.

Form context and options go in
The model scores every option
The executor picks the highest
Cua Driver sets a value or clicks

Good to know: The card says this checkpoint is not a formally mapped release of the cua-s1-form-v0 research profile. The jev-use chooser does not load it.

• License: MIT
• Cost to run: none named
• Data it reaches: the form elements and documents you pass in, and a live Cua Driver session if you use the planner loop
• Not for: generating text, or loading pickle copies of the model, which the card says to avoid

On Hugging Face, 13 November 2026

• Task: other
• Library: cua-s1
• Formats: Safetensors
• Downloads in the last 30 days: 0
• Likes: 119
• License: mit
• Languages: en
• Quantized versions: 2
• Tags: cua-s1, computer-use, form-filling, option-attention, system-one, safetensors, other

Links

• Model: https://huggingface.co/cua-ai/cua-s1-forms
• Files: https://huggingface.co/cua-ai/cua-s1-forms/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download cua-ai/cua-s1-forms

Then follow the usage in its model card. Nothing in this kit runs by itself.