A training-free layer that turns an open LLM into a decision model. It reads typed choice, yes/no and score answers from one prefill, removes option-order bias with no labels and, from a few hundred labels, calibrates or fits a closed-form head. Weights stay untouched.
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, score, noul, classify, route
choice, score, noul
Architecture
anyjev
cygnet
Fine-tuned from
qwen/qwen3-8b
google/gemma-4-12b-it
License
apache-2.0
mit
Availability
Open weights
Open weights
Hosted by
—
—
Input price
—
—
Decision accuracy
77.1%
87.9%
Calibration error
0.034
—
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 anyjev and cygnet?
anyjev is from Nokia Applied Research and cygnet from Blockbrain Labs. Both have open weights you can download and run. Both answer choice, score and noul questions. Only anyjev answers classify and route. anyjev is the smaller model, at 8.0B parameters to 12B. anyjev is licensed apache-2.0; cygnet, mit.
Which is more accurate, anyjev or cygnet?
They report on different suites — anyjev 77.1% on LocalLLaMA/typed-decisions (2,000 held-out decisions), L2 with 300 labels per question, 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, anyjev or cygnet?
anyjev: Free (open weights). cygnet: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run anyjev or cygnet locally?
Yes, both: systemone pull nokia-applied-research/anyjev and systemone pull blockbrain-labs/cygnet download the weights.
Evaluation suite
LocalLLaMA/typed-decisions (2,000 held-out decisions), L2 with 300 labels per question
