NeoBERT
modelA 250M-parameter next-generation bidirectional encoder pretrained from scratch, reaching state of the art on MTEB versus ModernBERT and RoBERTa. Heavily adopted (~39k HuggingFace downloads, 197 likes), from Sarath Chandar's lab. A first-in-series modern-encoder release modernizing the BERT recipe with contemporary architecture, data, and training length. Released under MIT.
Model Details
Architecture DENSE
Parameters 250M
License MIT