Introducing the Ettin Reranker Family

Embedding models encode queries and documents separately, relying on vector similarity, which misses the direct interaction between texts. Cross-encoders (rerankers) allow joint attention for higher accuracy but are too expensive to run over an entire corpus. The standard production pattern is retrieve-then-rerank: a fast embedder retrieves top-K candidates, then a cross-encoder re-orders them. The new Ettin Reranker family addresses the need for accurate, efficient rerankers at multiple sizes.

All six models are built on ModernBERT backbones from the Ettin encoder suite and trained via pointwise MSE distillation from the strong teacher mxbai-rerank-large-v2 on 143 million (query, document, score) triples. The architecture uses a modular Transformer that enables unpadded inputs and Flash Attention 2, yielding up to 8.3x speedup over fp32+SDPA baselines. The same single-stage training recipe, with only learning rate and batch size varying by model size, was used across all sizes from 17M to 1B parameters.

The smallest model (17M) outperforms legacy MiniLM rerankers while being faster, and the 1B model matches its teacher within 0.0001 NDCG@10 on MTEB while running 2.4x faster. The 150M model is the strongest under-600M reranker tested. All models accept up to 8K tokens and are released under Apache 2.0, along with training data and a complete ~150-line training script, enabling easy adoption and further improvement.