cua-s1-nano-0.1

cua-s1-nano-0.1

cua-s1-nano-0.1 is a small research classifier for closed-option computer-use decisions. It has about 855,000 trainable parameters and was trained from scratch. Given a screen state and a closed set of candidate (element, action) options, it scores every option in one parallel forward pass and selects the highest-scoring option for each element.

This checkpoint belongs to the Cua-S1 research family. It is not a general-purpose assistant. Do not treat a result on one task family as evidence of reliability outside it.

Files

Each checkpoint has 855,296 parameters. Tensors are stored only in safetensors, and the JSON sidecar holds the architecture config and a SHA-256 signature over the tensors. This repository contains no pickle files and does not redistribute the SigLIP backbone.

Intended use

• Research on closed-option computer-use GUI decisions (element and action selection) with a very small, fast architecture.
• Comparing a from-scratch specialist against larger checkpoints, such as cua-s1-4b-0.2, on cua-bench-s1.
• Studying task-family generalization, including out-of-domain checks.

Out of scope

• General-purpose or open-ended computer operation.
• Unsupervised operation on production accounts or sensitive data.
• Actions with financial, legal, medical, employment, safety, or other high-impact consequences.
• Bypassing access controls, consent, rate limits, or service policies.
• Treating a selected option as proof that the action is correct, safe, or succeeded.

How it works

An option-attention head uses each candidate option as a query against the element's context tokens and produces one logit per option. Options are always encoded as text by a shared byte-level option encoder. The context depends on the modality:

• Text: a small trainable byte-level transformer over a rendered accessibility-tree excerpt for the element. It needs no extra dependencies.
• Multimodal: a frozen vision backbone over a screenshot crop of the element, followed by a small trainable projection. This checkpoint's config selects siglip (google/siglip-base-patch16-224).

Scoring is one parallel forward pass, not autoregressive generation. Latency is in the low milliseconds per task on a GPU and under 100 ms on a CPU.

Training summary

The model was trained from scratch; it has no base model. It was trained on the data/v1 split of the cua-bench-s1 core GUI task families: form_filling, login_auth, consent_checkbox, multi_step_submit, pagination, and search_filter. cua-bench-s1 builds these families with its synthetic generator and its converters for AndroidControl (Apache-2.0) and GUI-360 (MIT). For exact provenance, see Data sources and provenance. The two-stage training recipe (base training, then an optional fine-tune) is libs/cua-s1/training/train_nano.py. No dataset is distributed with this checkpoint.

Evaluation

These are measured results from the cua-bench-s1 Results section:

• Held-out cross-dataset text split: task-level accuracy of 0.000 to 0.286, depending on family. The per-element text context carries no goal string.
• Same-distribution multimodal split (the split it was trained on): 1.000.
• chess, on the same 15-position subset used for every model: 0.000 task accuracy and 0.227 element accuracy in both modalities.
• game_control: 0.000 task accuracy and 0.333 element accuracy. Neither result survives chance correction.
• safety_gate: 0.000 zero-shot, and 1.000 after fine-tuning on that family's own split.
• External general_decision benchmark: no result, because the model cannot produce that benchmark's text-only decision shape.

The strong same-distribution result and weak cross-dataset result mean that this checkpoint mostly reflects its training distribution. It is not a general GUI decision model.

How to run

Pin the download to the revision the Cua-S1 documentation was verified against. The weights are unchanged at later revisions of this repository that change only documentation.

From a checkout of trycua/cua:

uv sync --project libs/cua-s1/python                          # text modality
uv sync --project libs/cua-s1/python --extra nano-vision     # adds the multimodal backbone
libs/cua-s1/python/.venv/bin/hf download cua-ai/cua-s1-nano-0.1 \
  --revision 1f93fd0fdcbe33740334948f967dff9f6c8e9f34 \
  --local-dir "$HOME/cua-s1-models/cua-s1-nano-0.1"

from pathlib import Path
from cua_s1.nano import load_nano_checkpoint, select_device

root = Path.home() / "cua-s1-models" / "cua-s1-nano-0.1"
# Validates the format, version, and SHA-256 tensor signature before returning.
model, collator, config = load_nano_checkpoint(root / "text", select_device("auto"))

The multimodal checkpoint loads the same way from root / "multimodal". On first use, it downloads the frozen google/siglip-base-patch16-224 backbone through Transformers.

jev-use chooser

The jev-use closed-candidate chooser supports only the Cua-S1-4B adapters. It fails setup for cua-s1-nano-0.1, and no Cua Driver integration runs this checkpoint yet. To drive the jev-use flow with a Cua-S1 model, use cua-s1-4b-0.2 and follow Get the weights and run inference. Use this checkpoint through cua_s1.nano directly, or through cua-bench-s1's model adapters.

Limitations

• It needs a closed, pre-enumerated option set for each element and cannot propose an action outside that set.
• It scores each element independently, so it cannot compare which element to act on next across a screen.
• Its small size trades capacity for speed.
• Multimodal behavior inherits the frozen backbone's limitations and depends on screenshot-crop quality.
• game_control and chess are out-of-domain checks only.
• No weights-backed CI or canonical Cua Driver desktop E2E covers this checkpoint.

Run it in an isolated environment with least-privilege credentials and bounded actions. Verify outcomes independently, and require human confirmation before consequential or irreversible actions.

License

The weights in this repository are licensed under Apache-2.0. The frozen google/siglip-base-patch16-224 backbone used by the multimodal checkpoint is not redistributed here: its own license governs it. The Cua-S1 source code is MIT-licensed in trycua/cua.