Post

This week we released ModernBERT, the first encoder to reach SOTA on most common benchmarks across language understanding, retrieval, and code, while running twice as fast as DeBERTaV3 on short context and three times faster than NomicBERT & GTE on long context.

ModernBERT is available to use today on Transformers (pip install from main). More details in our announcement post. huggingface.co/blog/modernb...

ModernBERT-base is the first encoder to beat DeBERTaV3-base on GLUE. ModernBERT is also competitive or top scoring on single vector retrieval, ColBERT retrieval, and programming benchmarks.

How did we do it? First, we brought all the modern LLM architectural improvements to encoders, including alternating global & local attention, RoPE, and GeGLU layers, and added full model unpadding using Flash Attention for maximum performance (illustrated in the next post).

Second, we carefully designed ModernBERT's architecture run to efficiently across most common GPUs. Many common older models don't consider the hardware they will run on and are slower than they should be. Not so with ModernBERT. (Full model sequence packing illustrated below)

Last, we trained ModernBERT on variety of data sources, including web docs, code, & scientific articles, for a total of 2 trillion tokens of English text & code. 1.7 trillion tokens at a short 1024 sequence length, followed by 300 billion tokens at a long 8192 sequence length.