jina-reranker-v3

Model Overview

jina-reranker-v3 is a 0.6B-parameter multilingual document reranker designed for listwise reranking. It employs a novel "last but not late interaction" architecture: instead of separate multi-vector encoding, the model performs causal self-attention between the query and all documents within a single context window, extracting relevance scores from the last token embedding of each document. This approach enables simultaneous processing of up to 64 documents within a 131K token context window.

Benchmark Performance

The model achieves state-of-the-art results on BEIR (61.94 nDCG@10) while being 10x smaller than generative listwise rerankers. The following table compares jina-reranker-v3 against other rerankers on standard benchmarks (nDCG@10):