NLLB-200 distilled 600M — GGUF for nllb.cpp
facebook/nllb-200-distilled-600M converted for
nllb.cpp: a C++ / ggml inference
engine. One file holds the weights and the
tokenizer, so nothing else is needed at runtime — no Python, no
tokenizer.json, no HF snapshot.
| file | size | use it for | sha256 |
|---|---|---|---|
nllb-600M-f16.gguf | 1.24 GB | the default. Greedy output identical to HF generate() on every golden | e8e202c28ae5f43a7905e4790f930e16807e14eeb294a844400e64be57606f9a |
nllb-600M-q8emb.gguf | 1.00 GB | smaller and ~1.33x faster on CPU; token_embd at Q8_0, rest f16. corpus chrF 96.57 vs HF, 8/9 goldens byte-identical | aebad3ca0507c8a882659e10cda4861ed246068ac871c91e784d97cd1b17a73e |
nllb-600M-f32.gguf | 2.47 GB | the reference precision: running the verification gates, and the input to nllb-quantize | 150f7eab66d6a7e469930a4ab66835277a0635868f6cc4d19b62723d6fe6aa5e |
More aggressive quantisations (whole-model Q8_0, Q5_K_M) were built and rejected on measured translation quality. They are not published on purpose.
License
The weights are Meta's NLLB-200, released under CC-BY-NC 4.0 — non-commercial use only. Converting them to GGUF does not change that. This repository redistributes them unmodified in value (f32), at reduced precision (f16) or partially quantised (q8emb), with attribution to:
NLLB Team et al., No Language Left Behind: Scaling Human-Centered Machine Translation, 2022. arXiv:2207.04672
The nllb.cpp code is licensed separately; see its repository.