aoiandroid/nllb200-coreml-1024-pal8

NLLB-200 Core ML 1024 — 8-bit weight-only palettized (pal8)

Core ML mlProgram packages derived from the FLOAT32 nllb200-coreml-1024 export, post-processed with 8-bit uniform palettization (coremltools.optimize.coreml.palettize_weights, weight-only). Runtime computation stays in float32; only stored weights are compressed (LUT), which avoids the decoder/KV numerical failures seen with full-graph FP16 passes.

This variant is separate from aoiandroid/nllb200-coreml-1024-palettized (different compression pipeline).

Base model (official)

Upstream weights and tokenizer follow facebook/nllb-200-distilled-600M licensing and use constraints. See that model card for intended use, limitations, and citation.

Item	Value
Max length	1024 (encoder and KV decoder)
Packages	Encoder, decoder init, decoder step + tokenizer + config

Files

Path	Description
NLLB_Encoder_1024.mlpackage	Encoder
NLLB_Decoder_1024_init.mlpackage	Decoder first step
NLLB_Decoder_1024_step.mlpackage	Decoder step with KV cache
tokenizer/	SentencePiece assets
config.json	Model config

App / model ID

on-device iOS client uses model identifier nllb200-coreml-1024-pal8 with NLLBCoreMLTranslationService.

Conversion

1. Start from FLOAT32 .mlpackage tree (nllb200-coreml-1024).
2. Run Notebooks/nllb/create_palettized_weight_only.py (uniform 8-bit palettize_weights).

Benchmark (Mac, Swift Testing)

15-line 30_text.txt subset to eng_Latn: see repo report Notebooks/nllb/output/NLLB_pal8_mac_translation_test_report.md.

License and intended use

Same as the base model: CC-BY-NC-4.0. Research and non-commercial use; not for production without your own compliance review.

Citation

@article{nllb2022,
  title={No Language Left Behind: Scaling Human-Centered Machine Translation},
  author={{NLLB Team} and others},
  journal={arXiv preprint arXiv:2207.04672},
  year={2022}
}