A 2.8 MB byte-level transformer for GUI form filling. For each form element it returns one probability per typed option (fill an entity, check, click or skip), using jevlike's option-attention head.
A Gemma 4 12B fine-tune for typed decisions that also chats and reads images. Ships as GGUF for llama.cpp, holds a 64K context on a 16 GB GPU, and serves /v1/systemone next to /v1/chat/completions.
Decides
choice
choice, score, noul, classify, route
Architecture
cua-s1
winnow
Fine-tuned from
—
google/gemma-4-12b-it
License
mit
apache-2.0
Availability
Open weights
Open weights
Hosted by
—
—
Input price
—
—
Decision accuracy
100.0%
85.7%
Calibration error
—
—
Valid action rate
—
—
Median latency
—
143 ms
p95 latency
—
—
Evaluation suite
196-decision evaluation on real forms
Figures are from each model’s manifest; accuracy and latency are what the publishers report, on their own suites and hardware. Add a third model.
Questions
What is the difference between cua-s1-forms and winnow?
cua-s1-forms is from Cua and winnow from EldanRing. Both have open weights you can download and run. Both answer choice questions. Only winnow answers score, noul, classify and route. cua-s1-forms is the smaller model, at 706K parameters to 12B. cua-s1-forms is licensed mit; winnow, apache-2.0.
Which is more accurate, cua-s1-forms or winnow?
They report on different suites — cua-s1-forms 100.0% on 196-decision evaluation on real forms, winnow 85.7% on JevBench public subset (231 items), Q8_0 — so the numbers do not rank them. Test both on your own labelled examples.
Which is cheaper, cua-s1-forms or winnow?
cua-s1-forms: Free (open weights). winnow: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run cua-s1-forms or winnow locally?
Yes, both: systemone pull cua/cua-s1-forms and systemone pull eldanring/winnow download the weights.
