Hugging Face releases Ettin Reranker family

hugging face released the ettin reranker family, six cross-encoder models ranging from 17m to 1b parameters built on modernbert architectures. developed using a distillation recipe on mixedbread ai scores, these models achieve state-of-the-art mteb retrieval performance at their specific sizes. they support long-context reranking up to 8k tokens and are optimized for flash attention 2 to deliver up to 8.3x throughput speedups. the release includes full training data and recipes for open reproduction.