aoiandroid/nllb200-coreml-256-ane-pal8

NLLB-200 distilled 600M — Core ML (seq 256, ANE, palettized / pal8)

Core ML .mlpackage bundles for encoder and KV-cache decoder (init + step), exported from facebook/nllb-200-distilled-600M. Fixed sequence length 256, 8-bit palettization (pal8) on weights for a smaller footprint and Apple Neural Engine–friendly deployment.

Contents

Artifact	Role
NLLB_Encoder_256.mlpackage	Encode source tokens
NLLB_Decoder_256_init.mlpackage	Decoder first step (no past KV)
NLLB_Decoder_256_step.mlpackage	Autoregressive step with past KV
config.json	Architecture metadata (seq_len, compute_units, etc.)
tokenizer/	Reference tokenizer files for the app pipeline

Companion variant

• Float32 weights (no palettization): aoiandroid/nllb200-coreml-256-ane

Config highlights (config.json)

• model_type: m2m_100 / M2M100ForConditionalGeneration
• seq_len: 256
• ane_optimized: true
• compute_units: CPU_AND_NE
• Weights in the .mlpackage bundles use pal8 palettization (see Hub file sizes vs the float32 repo).

Base model and license

• Source weights: NLLB-200 distilled 600M (Meta / FAIR).
• Original terms are CC-BY-NC 4.0. This repository only ships a derived Core ML export for on-device use; the same license restrictions apply.

Intended use

• macOS / iOS on-device translation via MLModel (or coremltools in Python for validation).
• Language direction is controlled in your app: tokenizer src_lang, forced target BOS (e.g. eng_Latn), and your beam/greedy loop around Decoder_256_step.
• Compare accuracy and latency against the float32 sibling on your devices; palettization can change numerical behavior.

Limitations

• Fixed max sequence length 256 at export time; longer inputs require truncation or a different export.
• Not a drop-in Hugging Face transformers checkpoint; this is a split Core ML deployment (encoder + decoder packages).

Maintainer

aoiandroid