NLLB-200
A research translation model covering 200 languages, built for low-resource ones.
NLLB-200 is a family of machine translation models from Meta AI's NLLB Team, presented in a July 2022 paper. It translates single sentences between 200 languages, with a strong focus on low-resource languages, and is released for research use along with the Flores-200 evaluation benchmark.
- Translates single sentences across 200 languages
- Uses a Sparsely Gated Mixture of Experts design
- Reported 44% BLEU gain over the previous state of the art
- Flores-200 benchmark covers over 40,000 translation directions
- Research model, not released for production deployment
What it is
NLLB-200 is a machine translation system built to handle 200 languages in one model, including many that earlier systems largely ignored. The NLLB Team at Meta AI published the work in July 2022. A 3.3 billion parameter checkpoint is among the released variants. Its main purpose is research, especially on low-resource languages, and it translates single sentences rather than whole documents. The team says the model reached a 44% BLEU improvement over the previous state of the art. Everything from the project, including models, data and the Flores-200 benchmark, was open sourced.
How it came about
The project began with interviews with native speakers to understand what low-resource communities actually need from translation. The paper appeared on arXiv on 11 July 2022, with a revised version in August 2022, listing 39 authors under the NLLB Team name including Marta R. Costa-jussà, James Cross and Guillaume Wenzek. Alongside the models, the team released Flores-200, a human-translated benchmark covering every language in the system, and shared training code through the Fairseq repository.
Inside the model
The model uses a Sparsely Gated Mixture of Experts design, a form of conditional compute where only part of the network activates for each token. Training combined parallel multilingual data from many sources with monolingual text drawn from Common Crawl, using new data mining methods aimed at low-resource languages. The team introduced architectural and training changes to fight overfitting across thousands of translation tasks. Text was split into sentences and processed with a SentencePiece tokenizer, released with the model. Inputs longer than 512 tokens were not part of training.
What it can do
NLLB-200 translates single sentences across 200 languages. The team evaluated more than 40,000 translation directions using Flores-200, reporting BLEU, spBLEU and chrF++ scores, plus human ratings through the XSTS protocol and a toxicity benchmark spanning all Flores-200 languages. They report a 44% BLEU gain over the prior state of the art. Limits include degraded quality on long sequences, no support for document translation, and untested use in specialised fields such as medical or legal text.
How to use it
Weights for variants including a 3.3B checkpoint are downloadable, with usage guidance in the Fairseq code repository. The model card states it is a research model, not released for production deployment, and its output cannot serve as certified translation. A SentencePiece tokenizer ships alongside the weights. The paper states that all contributions were open sourced.
Safety and testing
The team combined human evaluation with a toxicity benchmark covering every Flores-200 language. Their ethical notes flag risks: better translation could help education and information access in communities with many African languages, but could also expose people with lower digital literacy to misinformation or scams if misused. Training data mined from public web sources may still contain personally identifiable information despite cleaning, and residual mistranslations could affect decisions about health and safety.
Reception and debate
The model card frames NLLB-200 as aimed at the machine translation research community, with testing concentrated on the Wikimedia domain and limited investigation elsewhere. It warns that language variants within supported languages may not be captured and that users should make their own assessments.
Why it matters
By covering 200 languages and releasing Flores-200, the project set out groundwork for a universal translation system and gave researchers a benchmark with full language coverage. The open release of models, data and training code was intended to support further work on low-resource translation.
Questions people ask about NLLB-200
What is NLLB-200 for?
It is a research machine translation model covering 200 languages, built especially for low-resource languages. It translates single sentences between language pairs. The team says it is not meant for production deployment, document translation, or specialised domains such as medical and legal text.
Who built it and when?
The NLLB Team at Meta AI built it. The paper, No Language Left Behind: Scaling Human-Centered Machine Translation, was submitted to arXiv on 11 July 2022 with 39 listed authors, and revised in August 2022.
How large is it?
The model card provided describes a 3.3 billion parameter variant. The paper covers a family of models rather than a single size, and the exact training details are given in the paper itself.
How do I get it?
Weights, including the 3.3B checkpoint, are downloadable, and usage instructions sit in the Fairseq code repository. A SentencePiece tokenizer is released with the model. The team states that all contributions from the work were open sourced.
What are its limits?
Training used inputs up to 512 tokens, so longer sequences may lose quality. It handles single sentences, not documents, and was tested mainly on Wikimedia text. Output cannot be used as certified translation, and language variants may not be fully captured.
Good reading and sources
There is no English Wikipedia article on NLLB-200 yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2207.04672, huggingface.co/facebook/nllb-200-3.3B. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.