Currently supported machine translation engines:
1. Dedicated Translation Model
NLLB
NLLB Serve uses Facebook NLLB
2. Large Language Model
Gemma
LTEngine uses Google Gemma 3
Not Language Left Behind
Notes
- There are distilled and dense models, a LLB-200 1.3B parameters model distilled from NLLB-200 54B parameters model should give better results then LLB-200 1.3B parameters dense model.
References
Alternatives
MADLAD-400
Great Apache license from Google, MADLAD-400 top model is still not as good as NLLB top model at 54B parameters but its is much smaller at 10.7B parameters.
Opus-MT
High quality and great MIT license, with web app available.
Note this is not one big model for MANY languages, there is one model for EACH language pair.
Notes
- Problem with the Chinese-English models:
https://github.com/Helsinki-NLP/OPUS-MT-train/issues/13
they are available here
Tatoeba-Challenge/models at master 路 Helsinki-NLP/Tatoeba-Challenge 路 GitHub
in this model list, Chinese is "zho" and English is "eng"
Opennmt
Tools:
Lingva
If you must use Google Translate, try bypassing its API with this: