NLLB (No Language Left Behind)

Definition

NLLB (No Language Left Behind) is an open machine translation model from Meta AI designed to provide high-quality translation across 200 languages, with a particular focus on low-resource languages historically underserved by translation technology.

Overview

No Language Left Behind was released by Meta AI in 2022 as a research and engineering effort to close the translation quality gap for the roughly 4,000 languages spoken worldwide that lacked usable machine translation, many of them African, South Asian, and Indigenous languages. Rather than treating low-resource translation as an afterthought bolted onto a system optimized for high-resource pairs like English-French, the NLLB team built dedicated data pipelines, evaluation benchmarks, and modeling techniques specifically for the long tail of languages. The core NLLB model is a dense or sparsely-gated Mixture-of-Experts (MoE) Transformer trained on a massive mined parallel corpus, supplemented by the FLORES-200 evaluation benchmark that Meta created to measure translation quality consistently across all 200 supported languages, including many with no prior standardized benchmark. Meta also published LASER3, a language-agnostic sentence embedding model used to mine parallel sentence pairs from noisy web text for low-resource languages where clean parallel corpora don't naturally exist. NLLB models were released in multiple sizes, from distilled 600M and 1.3B parameter versions suitable for on-device or low-latency use, up to a 54.5B parameter MoE flagship model, all under an open license permitting research and commercial use. This let NLLB become a widely adopted backbone: Wikipedia's translation tooling, Meta's own products, and numerous third-party translation apps and research projects have built on NLLB rather than training comparable low-resource translation capability from scratch. NLLB is frequently cited alongside SeamlessM4T (which extends NLLB's text-translation lineage into speech) and Google's own massively multilingual models as one of the major open efforts to make translation technology work equitably across the world's languages rather than just the dozen or so with abundant training data.

Key Concepts

• Translates across 200 languages, with strong emphasis on low-resource languages
• Sparsely-gated Mixture-of-Experts architecture at its largest scale (54.5B parameters)
• FLORES-200 benchmark enables consistent evaluation across all supported languages
• LASER3 sentence embeddings used to mine parallel data for low-resource pairs
• Released in multiple sizes from 600M to 54.5B parameters
• Open license permitting research and commercial deployment
• Purpose-built data pipelines for languages with little to no digital parallel text
• Widely adopted as a backbone for third-party and open-source translation tools

Use Cases

Frequently Asked Questions

Related in AI Models