Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference

TL;DR

ModernBERT is a highly optimized encoder model that achieves state-of-the-art performance, efficiency, and long-context capabilities, making it suitable for diverse tasks and inference on standard GPUs.

Contribution

Introduces ModernBERT, a new encoder model with significant Pareto improvements in speed, memory, and long-context handling, trained on 2 trillion tokens.

Findings

State-of-the-art results on diverse classification tasks

Superior speed and memory efficiency for inference

Effective long-context processing up to 8192 tokens

Abstract

Encoder-only transformer models such as BERT offer a great performance-size tradeoff for retrieval and classification tasks with respect to larger decoder-only models. Despite being the workhorse of numerous production pipelines, there have been limited Pareto improvements to BERT since its release. In this paper, we introduce ModernBERT, bringing modern model optimizations to encoder-only models and representing a major Pareto improvement over older encoders. Trained on 2 trillion tokens with a native 8192 sequence length, ModernBERT models exhibit state-of-the-art results on a large pool of evaluations encompassing diverse classification tasks and both single and multi-vector retrieval on different domains (including code). In addition to strong downstream performance, ModernBERT is also the most speed and memory efficient encoder and is designed for inference on common GPUs.