“The Jemma models were trained on data from over 140 languages.”