Landmark multilingual model covering 101 languages (50+ previously underserved), built through a global collaboration of 3,000+ researchers from 119 countries. 13B parameters, fine-tuned on mT5. Published at ACL 2024.

Accompanied by the Aya Dataset: 204K human-annotated instruction pairs and 513M augmented instances across 114 languages. Apache 2.0. Pioneered the open, community-driven approach to multilingual AI that defined Cohere Labs' identity.

Model Details

Architecture DENSE
Parameters 13B

Paper

Venue ACL 2024
open-weightmultilingualresearch

Related