mT5: a multilingual T5 over 101 languages
In one sentence Google Research publishes mT5, a T5 variant pre-trained on mC4 (multilingual Common Crawl) over 101 languages, which becomes a standard baseline for many cross-lingual NLP tasks.
T5 was a Google model known for turning any NLP task into a "text in → text out" problem. But it was English-only. For anyone working in another language — Italian included — it wasn't usable.
mT5 is the multilingual version: the same model, but trained on 101 different languages, from Arabic to Vietnamese. It can translate, summarize, and answer questions across all of them, and learns specific tasks much faster than a monolingual model.
For anyone building applications that need to work in multiple countries without retraining a model for each one, it's a practical turning point.
Companies
Tools
mT5
Tags
Sources