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.
Typed decisions from Google's frozen Gemma-4-12B-it, with no fine-tuning. A small shim over unmodified vLLM letters the options, reads the model's own probability for each letter at one answer position and applies one calibration temperature; one output token per decision.
Decides
choice
choice, score, noul
Architecture
cua-s1
cygnet
Fine-tuned from
—
google/gemma-4-12b-it
License
mit
mit
Availability
Open weights
Open weights
Hosted by
—
—
Input price
—
—
Decision accuracy
100.0%
87.9%
Calibration error
—
—
Valid action rate
—
—
Median latency
—
50 ms
p95 latency
—
—
Evaluation suite
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 cygnet?
cua-s1-forms is from Cua and cygnet from Blockbrain Labs. Both have open weights you can download and run. Both answer choice questions. Only cygnet answers score and noul. cua-s1-forms is the smaller model, at 706K parameters to 12B.
Which is more accurate, cua-s1-forms or cygnet?
They report on different suites — cua-s1-forms 100.0% on 196-decision evaluation on real forms, cygnet 87.9% on JevBench public set (231 items), JevBench CLI — so the numbers do not rank them. Test both on your own labelled examples.
Which is cheaper, cua-s1-forms or cygnet?
cua-s1-forms: Free (open weights). cygnet: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run cua-s1-forms or cygnet locally?
Yes, both: systemone pull cua/cua-s1-forms and systemone pull blockbrain-labs/cygnet download the weights.
