Hugging Face releases six Ettin reranker models with distillation training recipe

Hugging Face published six new reranker models scaled from 17 million to 1 billion parameters, all built on the ModernBERT architecture from Johns Hopkins University's Ettin suite. The models are released as Sentence Transformers CrossEncoder implementations, integrable into production pipelines with minimal code.

Unlike embedding models, which encode query and document separately, rerankers perform joint encoding where query and document attend to each other across all transformer layers. This produces more accurate relevance scores but at higher computational cost, making them impractical to run over entire document corpora. The published models are intended for retrieve-then-rerank workflows: a fast embedding model retrieves top-K candidates, then the reranker re-orders those K with higher accuracy.

The six models were trained using knowledge distillation, with pointwise MSE loss from mixedbread-ai's mxbai-rerank-large-v2 as the teacher signal. Training data combined lightonai's embedding pre-training and fine-tuning datasets with additional reranking-specific curation. All models support up to 8,192 tokens of input context and achieve measured speedups of 1.7x to 8.3x when using flash attention and bfloat16 precision.

Hugging Face accompanied the release with a full training recipe, evaluation methodology, and dataset composition. The organization also added a new agent skill to Sentence Transformers v5.5.0 enabling developers to invoke AI coding assistants to fine-tune rerankers on custom data without manual engineering.

Performance is reported on MTEB English v2 retrieval tasks paired with various embedding models, with results detailed in the blog post for five embedder variants beyond the headline google/embedding-gemma-300m pairing.