KaLM-Reranker-V1: Fast but Not Late Interaction for Compressed Document Reranking

KaLM-Reranker-V1 is a new fast but not late-interaction (FBNL) document reranker that decouples query and passage computation using an encoder-decoder architecture. It features three model sizes (Nano: 0.27B, Small: 1B, Large: 4B parameters) and employs Matryoshka embedding pooling for efficient passage encoding, maintaining strong relevance modeling through cross-attention. Benchmark results on BEIR, MIRACL, and LMEB show that KaLM-Reranker-V1 achieves state-of-the-art performance, rivaling larger models while offering significant efficiency advantages, making it a valuable tool for practitioners in retrieval systems.