Jina-reranker-v3 is a 0.6B-parameter multilingual document reranking model that introduces a novel "last but not late interaction" approach. Unlike late interaction models like ColBERT, it performs causal self-attention on queries and documents within the same context window, enabling rich cross-document interactions before extracting contextual embeddings from the last token of each document. This compact architecture achieves state-of-the-art BEIR performance of 61.94 nDCG@10, while being 10x smaller than generative listwise reranking models.
