NLLB-200 and what low-resource language means

NLLB-200 is the model family from Meta's No Language Left Behind (NLLB) project, which translates directly between pairs of more than 200 languages, with an emphasis on low-resource ones. A low-resource language is one with little translated text available for training, and the term has no single numeric definition. For a program office weighing a released model for documents that cannot go to a cloud service, three questions matter: what low-resource means, how the scores were produced, and what the project's own pages say the model may be used for.

What low-resource means

A survey of low-resource machine translation by Haddow and colleagues notes that around 7,000 languages are spoken in the world and that almost all language pairs lack significant training resources. The authors treat resourcedness as a continuum in which any cutoff is arbitrary, and they note that the definition changes over time as data collection and methods improve. They also separate simulated low-resource settings, where a high-resource pair is deliberately restricted, from genuinely low-resource languages. The main constraint is parallel data, meaning sentences paired with their translations. Monolingual text is usually more plentiful.

The NLLB team sets thresholds for its own work in the Nature paper. It defines a high-resource language as one with at least 1 million sentences of aligned text with another language, and a very low-resource language as one with fewer than 100,000 samples across all pairings. The NLLB repository README names Asturian, Luganda, and Urdu as examples of low-resource languages the project covers. It states the aim of helping people communicate with anyone, anywhere, and of fostering more research in low-resource languages.

What was released

The README lists five models: a 54.5-billion-parameter mixture-of-experts model, dense models of 3.3 billion and 1.3 billion parameters, and distilled models of 1.3 billion and 600 million parameters. The README describes the distilled models as smaller distilled versions of NLLB-200. The model card for the 600M variant describes NLLB-200 as a machine translation model primarily intended for research, especially for low-resource languages.

Three data resources came with them, according to the Nature paper:

• FLORES-200, a human-translated evaluation benchmark. The Data availability section says it holds evaluation data in 204 languages, and the paper says it supports measurement across approximately 40,000 translation directions.
• NLLB-Seed, human-translated seed training data in 39 languages.
• Toxicity-200, wordlists for detecting toxicity in 200 languages.

The paper also releases SPM-200, a SentencePiece tokenizer covering 200 languages, to enable the spBLEU metric. The NLLB README and the FLORES repository are both archived, and the README banner gives March 20, 2026 as the archive date. The FLORES page says newer FLORES and NLLB-Seed versions are maintained by the Open Language Data Initiative.

What the evaluation numbers mean

The Nature paper reports that chrF++ was the primary automatic metric, computed with sacreBLEU settings, with BLEU and spBLEU also reported. The sacreBLEU documentation describes chrF++ as a revised chrF that takes word n-grams into account in addition to character n-grams.

Human evaluation used a protocol called XSTS, which uses a five-point scale that emphasizes meaning preservation over fluency, with 3 as the acceptability threshold. The paper evaluated a sample of 55 language directions, including 8 into English and 27 out of English. The model card also lists toxicity of generated translations among the measurements.

The headline result is an average 44 percent improvement over previous state-of-the-art systems. The abstract says this is measured by BLEU, and the body describes it as almost +7.3 spBLEU against the nearest state-of-the-art system. The same figure therefore appears under two metric names in one paper.

The model card also states a scope limit: the model was tested on the Wikimedia domain, with limited investigation of other domains. Documents outside the Wikimedia domain fall under that limited investigation.

License terms as the pages state them

The project's pages state the following.

• The README says NLLB code and fairseq are MIT-licensed, and that all models are licensed under CC-BY-NC 4.0, with the text in a Model LICENSE file. The README itself does not use the words "non-commercial."
• The Nature paper describes its resources as freely available for non-commercial use.
• The model card for the 600M distilled model gives "License: CC-BY-NC" and says NLLB-200 is a research model and is not released for production deployment.
• The FLORES repository's Licenses section lists FLORES-200 as CC-BY-SA 4.0.

The license text itself is in the files these pages point to. A program office should read that text rather than rely on a summary of it.

What to check

• Which license statement applies to the intended use: the README's CC-BY-NC 4.0, the model card's "License: CC-BY-NC," or the Nature paper's non-commercial wording?
• Does the intended use fit the model card's description of NLLB-200 as a research model not released for production deployment?
• Which definition of low-resource is the vendor using, and does the language fall under the NLLB team's threshold of fewer than 100,000 samples across all pairings?
• Which metric is quoted, chrF++, BLEU, or spBLEU, and with which settings?
• Was the evaluation run on text like the program's own documents, or on the Wikimedia domain?

Azcuna's BABEL reads, translates, and searches foreign-language documents entirely on-device, with no cloud and no connection, and keeps every claim tied to the passage it came from. The long read on offline machine translation goes further.

Sources

1. 1Scaling neural machine translation to 200 languages · Nature, volume 630, pages 841-846 (2024), NLLB Team; open-access copy on PubMed Central
2. 2Survey of Low-Resource Machine Translation · arXiv (Haddow, Bawden, Miceli Barone, Helcl, Birch)
3. 3fairseq, NLLB branch: No Language Left Behind (README) · Meta AI Research (facebookresearch on GitHub)
4. 4facebook/nllb-200-distilled-600M (model card) · Meta AI Research, hosted on Hugging Face
5. 5FLORES (repository) · Meta AI Research (facebookresearch on GitHub)
6. 6sacreBLEU project README · sacreBLEU project (Matt Post and contributors)