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 Thai-and-English System One model. Qwen3.5-0.8B continued-pretrained on about 5B Thai tokens, its language-model head replaced by a 256-way slot head (slot 255 abstains), trained on 2–3M decision examples and temperature-calibrated per question type.
Decides
choice, score, noul
choice, score, noul, classify, route
Architecture
laya
openthai
Fine-tuned from
convai-innovations/laya
qwen/qwen3.5-0.8b-base
License
apache-2.0
apache-2.0
Availability
Open weights
Open weights
Hosted by
—
—
Input price
—
—
Decision accuracy
—
74.3%
Calibration error
—
—
Valid action rate
—
—
Median latency
—
40 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 laya-mlx and openthai-systemone?
laya-mlx is from aac6fef and openthai-systemone from iApp Technology. Both have open weights you can download and run. Both answer choice, score and noul questions. Only openthai-systemone answers classify and route. laya-mlx is the smaller model, at 421M parameters to 800M.
Which is more accurate, laya-mlx or openthai-systemone?
Only openthai-systemone publishes an accuracy figure (74.3% on Bespoke public subsets (13, macro average)); laya-mlx does not, so there is no comparison to make without your own test.
Which is cheaper, laya-mlx or openthai-systemone?
laya-mlx: Free (open weights). openthai-systemone: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run laya-mlx or openthai-systemone locally?
Yes, both: systemone pull aac6fef/laya-mlx and systemone pull iapp-technology/openthai-systemone download the weights.
