← All models · mBERT language detection
use case · basics
The core move: give the model some text and it returns one of 201 languages with a confidence score, plus the full probability distribution over every language. Try the samples or type your own.
Each language has its own tell — the Cyrillic script for Russian, Han characters for Chinese, Thai script, or the function words and morphology of German and Spanish. Because the model reads script + morphology + function words in one shared embedding space, the model doesn't need a separate detector per language.