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.
Native MLX FP16 conversion of Laya's multilingual checkpoint (mmBERT) for Apple silicon, with a 1,024-token context. The base for multilingual fine-tunes in Laya Studio.
Decides
choice, score, noul
choice, score, noul
Architecture
cygnet
laya
Fine-tuned from
google/gemma-4-12b-it
convaiinnovations/laya-multilingual
License
mit
apache-2.0
Availability
Open weights
Open weights
Hosted by
—
—
Input price
—
—
Decision accuracy
87.9%
—
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 laya-multilingual-mlx?
cygnet is from Blockbrain Labs and laya-multilingual-mlx from aac6fef. Both have open weights you can download and run. Both answer choice, score and noul questions. cygnet reads up to 16K tokens of state, against 1K tokens for laya-multilingual-mlx. laya-multilingual-mlx is the smaller model, at 322M parameters to 12B. cygnet is licensed mit; laya-multilingual-mlx, apache-2.0.
Which is more accurate, cygnet or laya-multilingual-mlx?
Only cygnet publishes an accuracy figure (87.9% on JevBench public set (231 items), JevBench CLI); laya-multilingual-mlx does not, so there is no comparison to make without your own test.
Which is cheaper, cygnet or laya-multilingual-mlx?
cygnet: Free (open weights). laya-multilingual-mlx: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run cygnet or laya-multilingual-mlx locally?
Yes, both: systemone pull blockbrain-labs/cygnet and systemone pull aac6fef/laya-multilingual-mlx download the weights.
