M2M-100 is the translation model described in “Beyond English-Centric Multilingual Machine Translation,” posted to arXiv on October 21, 2020 by Angela Fan, Shruti Bhosale, Holger Schwenk, and colleagues at Facebook AI. It was presented as the first true many-to-many model able to translate directly between any pair of 100 languages rather than pivoting through English.
Most earlier multilingual systems were English-centric: to go from, say, Chinese to French they would translate Chinese to English and then English to French, which loses meaning at each hop. M2M-100 was trained on a large dataset that included thousands of non-English translation directions, built by mining parallel text and combining dense model scaling with language-specific sparse parameters. The paper reported gains of more than 10 BLEU points on direct, non-English translations compared with English-pivot baselines, while staying competitive with the best single-pair systems.
The work was a direct stepping stone to Meta’s later No Language Left Behind effort, which pushed the same many-to-many idea from 100 to 200 languages. M2M-100 established that broad direct translation was feasible and that the biggest improvements showed up exactly where prior systems were weakest, the non-English language pairs.
For organizations operating across many regions, the practical payoff is translating between any two supported languages directly, without the quality loss and added latency of bouncing everything through English.