Ettin Reranker Family: Six New Sentence‑Transformer Cross‑Encoders

Tom Aarsen has released six new Sentence‑Transformer Cross‑Encoder rerankers—ettin‑reranker‑17m‑v1, 32m‑v1, 68m‑v1, 150m‑v1, 400m‑v1, and 1b‑v1—built on Ettin ModernBERT encoders and trained with a pointwise MSE distillation recipe from mxbai‑rerank‑large‑v2. Each model accepts up to 8 K tokens of context, thanks to ModernBERT’s long‑context pre‑training, and can be used with a flash‑attention 2 implementation for higher throughput. The rerankers were evaluated on the MTEB (eng, v2) Retrieval benchmark, paired with six embedding models ranging from static‑retrieval‑mrl‑en‑v1 to jina‑embeddings‑v5‑text‑small‑retrieval, and achieved NDCG@10 scores that surpass retriever‑only baselines in many cases. The architecture consists of a 4‑module classification head on top of a plain AutoModel encoder, with CLS pooling and GeGLU layers, and the models are released under the Apache 2.0 license. Example usage shows that a simple CrossEncoder.predict or .rank call can rerank top‑K results with scores indicating relevance, and switching to a larger model trades speed for quality. The release also includes a full training script that can be run with the Sentence‑Transformers v5.5.0 train‑sentence‑transformers agent skill. By integrating these rerankers into a retrieve‑then‑rerank pipeline, developers can achieve higher ranking accuracy without a full cross‑encoder pass over the entire corpus.