aoiandroid/nllb200-coreml-128-mac-palettized

NLLB-200 Core ML 128 (Mac, Palettized)

Core ML conversion of NLLB-200 (No Language Left Behind, ~600M parameters) for macOS, with palettized weights to reduce memory and improve on-device performance. Fixed sequence length 128 tokens (short text / chat).

• Base: facebook/nllb-200-distilled-600M
• Variant: 128-token encoder/decoder, palettized, Mac-optimized
• Use case: Short sentences, chat, ~80–100 words per segment

Contents

• NLLB_Encoder_128.mlpackage – NLLB encoder (input_ids, attention_mask → hidden states)
• NLLB_Decoder_128.mlpackage – NLLB decoder (input_ids, encoder_hidden_states, encoder_attention_mask → logits)
• tokenizer/ – Tokenizer (tokenizer.json, tokenizer_config.json, sentencepiece)
• config.json – Model config

Device

Mac only. For iPhone use nllb200-coreml-128-iphone-palettized; for iPad use nllb200-coreml-128-ipad-palettized.

Usage (macOS / TranslateBlue)

1. Download this repo (e.g. via Hugging Face Hub or TranslateBlue in-app download).
2. Load NLLB_Encoder_128.mlpackage and NLLB_Decoder_128.mlpackage with Core ML; tokenize with the included tokenizer.
3. Run encoder then decoder in a loop (argmax next token until EOS or max length 128).

Source/target language is controlled by the NLLB-200 language code prefix (e.g. eng_Latn, jpn_Jpan).

Related repos

Device	128 tokens	256 tokens
iPhone	nllb200-coreml-128-iphone-palettized	nllb200-coreml-256-iphone-palettized
iPad	nllb200-coreml-128-ipad-palettized	nllb200-coreml-256-ipad-palettized
Mac	this repo	nllb200-coreml-256-mac-palettized

License

CC-BY-NC-4.0 (inherited from NLLB-200).