NLLB-200 distilled 600M — Core ML KV (seq 256, palettized / pal8)
Core ML packages for encoder and KV-cache decoder (init + step) exported from facebook/nllb-200-distilled-600M, sequence length 256, 8-bit palettization (pal8) for Apple Neural Engine–friendly deployment.
Contents
NLLB_Encoder_256.mlpackageNLLB_Decoder_256_init.mlpackageNLLB_Decoder_256_step.mlpackageconfig.json, bundled tokenizer/ (reference)
Base model
- facebook/nllb-200-distilled-600M
- Original NLLB-200 licensing applies (CC-BY-NC). This repo only redistributes a derived Core ML export for on-device use.
Usage notes
- Intended for macOS / iOS inference via
MLModel (or Python coremltools for tests).
- Source language is selected via tokenizer /
src_lang; target via forced BOS (e.g. eng_Latn) in your app pipeline.
- Quality and parity should be validated on your target devices (ANE vs CPU/GPU paths may differ).
Organization
