Jina-reranker-v3 is a multilingual listwise reranker with 0.6B parameters that introduces a "last but not late" interaction approach. Unlike late interaction models like ColBERT, it applies causal attention between the query and all candidate documents, enabling rich interactions before extracting contextual embeddings from the final tokens of each document. This model achieves a state-of-the-art BEIR performance of 61.94 nDCG@10, while remaining significantly smaller than other models with similar performance.
