Meta open-sources No Language Left Behind translation model

Meta AI open-sourced NLLB-200, a single model able to translate between 200 languages, as part of its No Language Left Behind project. The model’s stated focus was low-resource languages that mainstream translation tools had largely ignored: Meta said NLLB-200 supported 55 African languages with quality results, against fewer than 25 covered by widely used commercial systems at the time.

The release also included FLORES-200, a many-to-many evaluation dataset that let researchers benchmark translation quality across roughly 40,000 language-direction pairs — most of which had no prior standard evaluation set at all. Meta reported NLLB-200 beat the previous state of the art by an average of 44% across the benchmarks it tested, and published the model weights, training code and evaluation tools rather than keeping the system internal or API-only. A technical paper detailing the training data curation, model architecture and evaluation methodology followed the same week.

The project’s significance sits in the open-vs-closed argument that ran through the period: translation for widely-spoken languages such as English, Mandarin or Spanish was already commercially well served, but low-resource languages had comparatively little economic incentive behind them, and Meta’s decision to open the weights meant researchers and organisations working in those languages could build on the model directly rather than wait for a commercial product to prioritise their language.