Native MLX FP16 conversion of Laya for Apple silicon. The same ModernBERT-large encoder, decision transformer and heads, running in MLX with no PyTorch; the checkpoint Laya Studio fine-tunes from.
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, score, noul
choice, score, noul, classify, route
Architecture
laya
winnow
Fine-tuned from
convai-innovations/laya
google/gemma-4-12b-it
License
apache-2.0
apache-2.0
Availability
Open weights
Open weights
Hosted by
—
—
Input price
—
—
Decision accuracy
—
85.7%
Calibration error
—
—
Valid action rate
—
—
Median latency
—
143 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 laya-mlx and winnow?
laya-mlx is from aac6fef and winnow from EldanRing. Both have open weights you can download and run. Both answer choice, score and noul questions. Only winnow answers classify and route. winnow reads up to 64K tokens of state, against 512 tokens for laya-mlx. laya-mlx is the smaller model, at 421M parameters to 12B.
Which is more accurate, laya-mlx or winnow?
Only winnow publishes an accuracy figure (85.7% on JevBench public subset (231 items), Q8_0); laya-mlx does not, so there is no comparison to make without your own test.
Which is cheaper, laya-mlx or winnow?
laya-mlx: Free (open weights). winnow: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run laya-mlx or winnow locally?
Yes, both: systemone pull aac6fef/laya-mlx and systemone pull eldanring/winnow download the weights.
