Meet mmBERT: An Encoder-only Language Model Pretrained on 3T Tokens of Multilingual Text in over 1800 Languages and 2–4× Faster than Previous Models