Ettin Reranker: 6 Models Released

Hugging Face released 6 Sentence Transformers CrossEncoder rerankers based on Ettin ModernBERT. Sizes range across 6 tiers from 17M to 1B, and each is claimed to be SOTA at its scale.

• Training was done via pointwise MSE distillation.
• The teacher model was mixedbread-ai/mxbai-rerank-large-v2.
• The data used was cross-encoder/ettin-reranker-v1-data, created by mixing portions of lightonai/embeddings-pre-training and lightonai/embeddings-fine-tuning.

In practical use, they can be plugged in directly via sentence_transformers.CrossEncoder, and they rerank top-K candidates as part of a retrieve-then-rerank pipeline. All models support up to 8K tokens, and using bfloat16 and flash_attention_2 reportedly delivers a 1.7x–8.3x speedup over the default.

The architecture is a ModernBERT-family backbone with a CLS pooling + Dense + LayerNorm head on top. Instead of AutoModelForSequenceClassification, AutoModel is used to take advantage of sequence unpadding and Flash Attention 2, and the authors noted that CLS pooling performed better than mean pooling. MTEB (eng, v2) Retrieval results were also released alongside.