Ettin Reranker family beats much larger models from 17M parameters up

The Sentence Transformers team released the Ettin Reranker family on May 19, six cross-encoder models that score how well a document answers a query and slot into the retrieve-then-rerank pipeline behind most modern search systems. The sizes range from 17M to 1B parameters, all Apache 2.0, all distilled from mixedbread's 1.54B reranker.

The interesting part is the quality-to-size ratio. The 17M model beats ms-marco-MiniLM-L12-v2, the legacy default in many retrieval stacks, by 0.05 NDCG@10 on MTEB while running 2x faster, and the 32M model beats BAAI's 568M bge-reranker-v2-m3 at 17 times fewer parameters. At the top end, the 1B variant matches its 1.54B teacher within 0.0001 NDCG@10 while being 54% smaller. All six handle up to 8,192 tokens of context, enough to rerank long documents without chunking, and the speed comes from running ModernBERT with Flash Attention 2 on unpadded inputs.

The team also published the 143M training triples and a 150-line training script, so you can fine-tune any size on your own retrieval data. Full benchmarks and a usage example are in the release post.

Why it matters

If your RAG stack still leans on an old MiniLM reranker, the 17M and 32M Ettin models are likely drop-in replacements that improve relevance and throughput on the same hardware. The 8K context also removes a common chunking workaround for long-document search.