aoiandroid/nllb200-coreml-1024

NLLB-200 CoreML 1024 (Float16)

CoreML conversion of facebook/nllb-200-distilled-600M (NLLB-200 / M2M100, distilled 600M parameters, 200 languages) for on-device inference on Apple platforms. Max sequence length 1024 matches the official model max_position_embeddings.

Base model (official)

This CoreML model is derived from facebook/nllb-200-distilled-600M (NLLB-200 distilled 600M on the Hub). Key official specs:

Item	Official value
Model type	M2M100 (encoder-decoder)
Parameters	~600M (distilled)
Languages	200 (Flores-200 coverage)
Max position embeddings	1024 (config.json)
Vocab size	256,206
License	CC-BY-NC-4.0

From the base model card: research and non-commercial use; single-sentence translation; not for production, medical, or legal domain; training used input lengths not exceeding 512 tokens (longer sequences may degrade). Not for certified translation. For full intended use, limitations, metrics (BLEU, spBLEU, chrF++), and ethical considerations, see the official model page.

Note: Other NLLB variants (e.g. larger or MoE) may support longer context (e.g. 128k); this CoreML conversion follows the distilled 600M config, which has max 1024.

Model description

• Base model: facebook/nllb-200-distilled-600M
• Precision: Float16
• Max length: 1024 tokens (encoder and decoder), aligned to official max_position_embeddings
• Architecture: Encoder + KV-cache decoder (init + step), compatible with variable sequence length 1..1024.

Files

File / folder	Description
NLLB_Encoder_1024.mlpackage	Encoder (source encoding)
NLLB_Decoder_1024_init.mlpackage	Decoder first step (with encoder outputs)
NLLB_Decoder_1024_step.mlpackage	Decoder subsequent steps (with past KV cache)
tokenizer/	SentencePiece tokenizer (same as base model)
config.json	Model config

Usage

1. Download this repo (or clone via Hugging Face Hub).
2. Load the three CoreML packages: encoder, decoder init, decoder step.
3. Encode source with the included tokenizer (use source/target language codes, e.g. eng_Latn, jpn_Jpan).
4. Run decoder in a loop: call init once with encoder outputs, then step with the last token and updated past key/value tensors until EOS or max length.

For a smaller, 8-bit quantized variant (same 1024 length), see aoiandroid/nllb200-coreml-1024-palettized.

Conversion

Produced with the notebook [nllb200_coreml_colab_v3_fixed.ipynb](local iOS project (private)) (CoreML encoder + KV decoder trace and export). Alignment: 1024 matches the official config.json of the base model.

License and intended use

• License: CC-BY-NC-4.0 (same as the base model).
• Intended use: Research and non-commercial use. Not for production deployment. General domain only (not for medical or legal domain). See the base model card for full intended use, limitations, and ethical considerations.

Citation

If you use this model, please cite the NLLB-200 paper and the base model:

@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}
}