aac6fef ports Laya decision encoder to MLX for Apple silicon

aac6fef has published laya-mlx, a native MLX FP16 conversion of the convaiinnovations/laya decision encoder built to run on Apple silicon without PyTorch or Transformers installed.

The model card describes it as a 'bidirectional decision encoder' on ModernBERT-large that supports 'choice, ordinal score, and boolean noul questions' inside a 512-token total context. The worked example is customer-service request routing: an agent is asked which department should handle a duplicate-billing complaint, followed by a boolean 'Does the customer ask for money back?'

The port ships at 843 MB, 0.4B parameters, in FP16 safetensors. On an Apple M3 Max with a 40-core GPU and 128 GB unified memory, '63/63 decision distributions match upstream PyTorch MPS FP32 on argmax', with a max probability difference of 0.0054443. Parameter names were adapted for MLX with 'no retraining or further quantization', per the card. Two of the researchers we track shared the page. License is Apache-2.0.