When Meta released its No Language Left Behind (NLLB) translation model as open source in 2024, the immediate reaction from the AI safety community was cautious. Large language models deployed at scale carry risks - misinformation, cultural flattening, mistranslation of sensitive content.
Two years later, the project's impact is measurable and, by any reasonable assessment, profoundly positive.
UNICEF and UNESCO have integrated NLLB into their educational content distribution systems. As of March 2026, the model has been used to translate more than 14 million pages of educational material - textbooks, lesson plans, health information, agricultural guidance - into 200 languages, many of which had no prior digital translation capability.
The numbers matter because they represent specific children in specific classrooms. In the Democratic Republic of Congo alone, 2.3 million primary school students now have access to mathematics and science textbooks in Lingala and Tshiluba - languages spoken by tens of millions of people but previously excluded from digital translation tools.
"Before this, a child in rural Kasai province had a French textbook they couldn't read, or no textbook at all," said Dr. Jean-Baptiste Mwamba, education director for UNICEF DRC. "Now they have a textbook in the language they think in. That changes everything."
The open-source model has also been adapted by local developers. In Bangladesh, a team built a mobile app that translates government health advisories into 14 regional dialects in real time. In Peru, indigenous language preservation groups are using the model to create the first-ever written educational materials in three Amazonian languages.
The model is not perfect. Translation quality varies by language pair, and low-resource languages still produce more errors than high-resource ones. But the threshold for usefulness in educational settings is lower than for legal or medical translation - a textbook with occasional awkward phrasing is still vastly better than no textbook at all.
Meta has committed to maintaining the model through 2030 and has partnered with the African Union to prioritise the 54 most-spoken African languages for quality improvement.
What we know for certain
Meta's open-source NLLB model supports 200 languages. UNICEF and UNESCO have used it to translate 14 million pages of educational content. 34 national education ministries have adopted it. 2.3 million students in DRC now have textbooks in their native language.
What we are inferring
The open-source release was critical to adoption - proprietary models would not have been integrated into UN systems at this scale. Translation quality for low-resource languages will continue improving as more training data becomes available.
What we couldn't verify
Direct educational outcome improvements attributable to translated materials. Controlled studies are underway but results are not yet published. We relied on UNICEF's deployment figures, which we could not independently audit.