Hugging Face Transformers to include ModernBERT as BERT replacement with 8k context in v4.48.0

AI Impact Summary

ModernBERT introduces an encoder-only family as a drop-in replacement for BERT, delivering 8k sequence length and faster processing with base (149M) and large (395M) variants. It can be used with standard MLM workflows via AutoModelForMaskedLM or the fill-mask pipeline, simplifying migration for existing BERT deployments. The feature set will be available in Hugging Face Transformers v4.48.0, with installation from main prior to release and optional use of Flash Attention 2 for higher efficiency on supported GPUs.