Introducing the Ettin Reranker Family

Hugging Face Blog · 1 month ago

The author released six reranker models built on ModernBERT encoders, ranging from 17 million to 1 billion parameters, trained via distillation from a larger teacher model. The smallest 17M model outperforms the 33M ms-marco-MiniLM-L12-v2 baseline by 0.051 NDCG@10 on MTEB while using half the parameters. Users can now swap in these rerankers as drop-in replacements in retrieve-then-rerank pipelines with improved quality and configurable speed-accuracy tradeoffs.