NLLB-200: Machine Translation Built for 200 Languages

The NLLB-200 research model translates among 200 languages. What its 2022 and 2024 papers measured, how the 44 percent gain is defined, and the limits.

Guide: AI Tools and IndustryLast verified: October 20265 min read

The most widely documented AI translator project of 2022 is No Language Left Behind (NLLB-200), a research model that translates among 200 languages. Its authors posted the paper on July 11, 2022 and published a peer-reviewed version in Nature on June 5, 2024. They reported an average 44 percent gain in BLEU score over the previous best systems, measured on a benchmark of more than 40,000 translation directions.

Key takeaways

  • NLLB-200 targets low-resource languages, which most earlier translation research skipped because little paired text exists for them.
  • The reported 44 percent gain is relative and averaged across many directions; the Nature text also gives about 7.3 spBLEU points as the absolute average gain.
  • The authors built a word-list toxicity detector for all 200 languages to flag harmful content added by a translation.
  • Human evaluation covered 55 language directions, far fewer than the automatic benchmark.
  • Models and code were released for non-commercial use.

What problem the project set out to solve

Most machine translation research has concentrated on a few dozen languages that have large collections of translated text. The arXiv paper states the aim in its title and abstract: scaling human-centered machine translation to 200 languages while keeping the output safe, high quality, and ethically considered. The authors began with interviews with native speakers of low-resource languages to understand what such communities want from the technology, and then built data sets and models.

The Nature version, "Scaling neural machine translation to 200 languages", defines the terms. A high-resource language has at least one million aligned sentences with another language. A very low-resource language has fewer than 100,000 samples across all its pairings in the data set. The project focuses on languages at the low end of that scale, where standard training recipes have little paired text to learn from.

How the system was built and tested

The abstract describes a Sparsely Gated Mixture of Experts model, trained on data gathered with mining techniques designed for low-resource languages, plus changes to the architecture and training to reduce overfitting. Evaluation used Flores-200, a benchmark of human-translated sentences, across more than 40,000 translation directions. The authors combined automatic scores with human evaluation and with a new toxicity benchmark covering every Flores-200 language.

The toxicity work addresses a specific failure: a translation that adds offensive words missing from the source. The authors built a detector named ETOX from word lists, called Toxicity-200, which were based on professional translations of English items and adapted by linguists. In their experiments on 10 directions, filtering misaligned training pairs for added toxicity removed about 30 percent of the sentences, reduced added toxicity, and raised the chrF++ score by about 5 percent.

Reading the 44 percent figure

The headline number is a relative improvement in BLEU over the previous state of the art, averaged across directions. The main text of the Nature paper adds absolute gains of about 7.3 spBLEU points across the full set and 7.0 points on an 87-language subset compared with the earlier FLORES-101 evaluation. The two statements describe the same result in different units.

An illustrative example shows why both matter. A score that rises from 10 to 14.4 has gained 44 percent in relative terms and 4.4 points in absolute terms. The relative figure looks larger when starting scores are low, and starting scores are low for exactly the languages this project targets.

Limits the authors state or imply

QuestionWhat the sources say
How many directions got human review?55 language directions, per the Nature text
Is the toxicity detector complete?No. Word lists are an imperfect proxy because toxicity is culturally sensitive and subjective
Are other limits documented?The paper says NLLB-200 still has limitations, documented in its supplementary information
Can the model be used commercially?Release is described as free for non-commercial use
Do scores prove a translation is safe to publish?No. Benchmark scores compare systems and do not certify a given text

A benchmark built from translated sentences also leaves out whole kinds of text, such as legal notices, medical instructions, and local idiom. A single research result for 200 languages does not show that every language works equally well, and an average hides differences between pairs.

What this means for someone choosing a translation tool

  • Check whether the language pair is in the tool's supported list, and when that list was last updated.
  • For health, legal, or safety text, have a fluent reader of both languages review the result.
  • Keep personal or confidential text out of tools whose data handling is unclear.
  • Treat a published average as a statement about a benchmark, not about a specific document.

This article describes one research project and does not recommend or compare commercial translation products. Product features and language lists change, and no claim here covers any app built on top of the released models.

FAQ

How many languages does NLLB-200 cover?

The model covers 200 languages, according to both the 2022 arXiv paper and the 2024 Nature paper.

Is the 44 percent improvement a measure of accuracy?

No. It is a relative improvement in BLEU, an automatic score that compares machine output with reference translations. It does not state what share of sentences is correct.

What counts as a low-resource language?

In the Nature paper, a very low-resource language has fewer than 100,000 samples across all its pairings in the data set, while a high-resource language has at least one million aligned sentences with another language.

Can the model be used in a company product?

The release is described as freely available for non-commercial use, so commercial use would need to be checked against the license terms published with the code.

Last verified: October 2026

Sources

  • No Language Left Behind: Scaling Human-Centered Machine Translation, NLLB Team, arXiv 2207.04672, 2022
  • Scaling neural machine translation to 200 languages, NLLB Team, Nature 630, 841-846, June 5, 2024

New to artificial intelligence?

Start with the basics, then follow the guides to tools, rules, and effects on society.

Start with AI Basics