aoiandroid/nllb200-coreml-1024-palettized

NLLB-200 CoreML 1024 (8-bit Palettized)

8-bit palettized CoreML variant of facebook/nllb-200-distilled-600M for on-device inference on Apple platforms (iPhone, iPad, Mac). Same architecture and max length (1024) as the float16 base, with smaller size and one build for all devices.

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: 8-bit palettized (encoder and decoder)
• Max length: 1024 tokens, aligned to official max_position_embeddings
• Architecture: Encoder + KV-cache decoder (init + step), same as aoiandroid/nllb200-coreml-1024.

Difference from the float16 repo

	nllb200-coreml-1024	nllb200-coreml-1024-palettized (this repo)
Precision	Float16	8-bit palettized
Size	Larger	Smaller
Use case	Higher precision on device	Same quality with lower memory and storage

Same file layout and usage (encoder + decoder init + decoder step + tokenizer + config).

Files

File / folder	Description
NLLB_Encoder_1024.mlpackage	Encoder (8-bit palettized)
NLLB_Decoder_1024_init.mlpackage	Decoder first step
NLLB_Decoder_1024_step.mlpackage	Decoder subsequent steps (KV cache)
tokenizer/	SentencePiece tokenizer
config.json	Model config

Usage

Same as the float16 model: load encoder, decoder init, and decoder step; encode source with the included tokenizer; run decoder loop with KV cache until EOS or max length. See aoiandroid/nllb200-coreml-1024 for a detailed usage description.

Conversion

Produced from the float16 CoreML export by applying 8-bit palettization (coremltools) in [nllb200_coreml_colab_v3_fixed.ipynb](local iOS project (private)). One palettized variant is built for all Apple devices.

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. See the base model card for full details.

Citation

See aoiandroid/nllb200-coreml-1024 for the NLLB-200 citation.