Researchers · Interviewed at COLM 2026
Cohere Labs

COLM 2026 · October 13, 2026 · 13 min
Does a model become multilingual just by getting bigger? Diana Abagyan from Cohere Labs argues it doesn't. Multilinguality has to be designed into every stage, from the tokenizer to post-training.
Recorded at COLM 2026 in San Francisco, this conversation covers Tiny Aya, Cohere Labs' family of 3B-parameter multilingual models. They cover 70 languages and are small enough to run on a phone, which matters for the communities these languages belong to.