Modelpublic

translate-studio/MADLAD400-3B-MT-int6-MLX

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes54downloads

Model Card

MADLAD-400 3B MT — int6 (MLX)

A 6-bit, group-size-64 MLX quantization of `google/madlad400-3b-mt` (a T5 encoder–decoder), for fast on-device translation on Apple Silicon. 400+ languages. Architecture unchanged — this repo only re-quantizes the original weights.


Base model	`google/madlad400-3b-mt`
Quantization	6-bit, group size 64 (the 32×16 relative-attention-bias tables are kept in full precision)
Format	MLX `safetensors` (+ SentencePiece tokenizer)
Size	~2.6 GB

License & attribution

Apache-2.0, inherited from the base model. All credit for the model goes to Google — see MADLAD-400. This repository only provides an MLX-quantized copy of the released weights.