NLLB-200
The facts below are what the maker publishes about NLLB-200, checked and dated. Our editors have not written about it yet.
The facts
Each one read from NLLB-200's own pages, with the date it was checked.
What it does
- Certified translations
- The model card says NLLB-200 translations cannot be used as certified translations.checked 12 Oct 2026
- Code license
- The repository says NLLB code and fairseq are MIT-licensed.checked 13 Oct 2026
- Data and accuracy caveats
- The model card warns that personally identifiable information may not be entirely eliminated from training data and that mistranslations may remain.checked 12 Oct 2026
- Document translation
- Nochecked 22 Sep 2026
- Evaluation
- Meta says it created the FLORES-200 dataset to evaluate translation quality across the supported languages.checked 13 Oct 2026
- Input length
- The model card says training inputs did not exceed 512 tokens and longer sequences may have lower translation quality.checked 12 Oct 2026
- Integration
- Dense models are available through Hugging Face Hub and supported by the Transformers library.checked 13 Oct 2026
- Intended users
- Meta presents the released models and research tools as resources for researchers and communities building or improving translation tools.checked 13 Oct 2026
- Language coverage
- Meta says the model includes 150 low-resource languages.checked 12 Oct 2026
- Library support
- The 3.3B model card provides usage instructions for the Transformers library and links to Google Colab and Kaggle notebooks.checked 12 Oct 2026
- License
- The 3.3B model card lists the license as CC-BY-NC-4.0.checked 12 Oct 2026
- Model license
- The repository states that the NLLB models are licensed under CC-BY-NC 4.0.checked 13 Oct 2026
- Model variants
- The repository lists NLLB-200 MoE at 54.5B parameters and dense versions at 3.3B and 1.3B, plus distilled versions at 1.3B and 600M.checked 13 Oct 2026
- Open-source release
- Meta says it open-sourced NLLB-200 models, FLORES-200, model training code, and code for recreating the training dataset.checked 12 Oct 2026
- Purpose
- NLLB-200 is an open-source AI project with models for direct translation between 200 languages, including low-resource languages.checked 13 Oct 2026
- Reported quality
- Meta reports that NLLB-200 exceeded the previous state of the art by an average of 44% in its evaluations.checked 13 Oct 2026
- Repository status
- The fairseq repository was archived on March 20, 2026, and is read-only.checked 13 Oct 2026
- Research resources
- The project releases model training and inference code, evaluation benchmarks, language identification models, LASER3 encoders, and data mining code.checked 13 Oct 2026
- Research use
- The model card describes NLLB-200 as primarily intended for machine translation research, especially for low-resource languages.checked 12 Oct 2026
- Safety evaluation
- Meta describes evaluating translations against human translations and filtering profanity and other offensive content using toxicity lists for supported languages.checked 13 Oct 2026
- Translation directions
- The project says its models support more than 200 × 200 translation directions.checked 13 Oct 2026
- Translation quality
- Meta reports an average 44% BLEU improvement over the previous state of the art.checked 12 Oct 2026
The company
- Founded
- 2004checked 28 Sep 2026
- Headquarters
- Menlo Park, California, United Stateschecked 28 Sep 2026
Alternatives to NLLB-200
- 1
locize
Free planAndroidiOSWebAPI8.9From $7 a month - 2
Crowdin
Free planAndroidiOSWebAPI8.6Free plan - 3
iTranslate
Free planAndroidiOSWebAPI8.6Free plan - 4
SimpleLocalize
Free planAndroidiOSWebAPI8.6Free plan - 5
XTM Cloud
AndroidiOSWebAPI8.5From $16,500 a year - 6
TranslatePress
Free planOpen sourceWebAPISelf-hosted8.5From €99 a year