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.
A 144M-parameter System One model on the multilingual mmBERT-small encoder. Takes a state, a question and 2 to 20 options and returns one decision with probabilities for choice, score and yes/no questions, on a CPU.
Decides
choice, score, noul
choice, score, noul, classify, route
Architecture
cygnet
julia
Fine-tuned from
google/gemma-4-12b-it
jhu-clsp/mmbert-small
License
mit
apache-2.0
Availability
Open weights
Open weights
Hosted by
—
—
Input price
—
—
Decision accuracy
87.9%
73.2%
Calibration error
—
—
Valid action rate
—
—
Median latency
50 ms
—
p95 latency
—
—
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 cygnet and julia-1?
cygnet is from Blockbrain Labs and julia-1 from Supersonic Labs. Both have open weights you can download and run. Both answer choice, score and noul questions. Only julia-1 answers classify and route. cygnet reads up to 16K tokens of state, against 8K tokens for julia-1. julia-1 is the smaller model, at 144M parameters to 12B. cygnet is licensed mit; julia-1, apache-2.0.
Which is more accurate, cygnet or julia-1?
They report on different suites — cygnet 87.9% on JevBench public set (231 items), JevBench CLI, julia-1 73.2% on typed-decisions test set (400 cases, 2,000 questions) — so the numbers do not rank them. Test both on your own labelled examples.
Which is cheaper, cygnet or julia-1?
cygnet: Free (open weights). julia-1: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run cygnet or julia-1 locally?
Yes, both: systemone pull blockbrain-labs/cygnet and systemone pull supersonic-labs/julia-1 download the weights.
Evaluation suite
JevBench public set (231 items), JevBench CLI
typed-decisions test set (400 cases, 2,000 questions)
