Jina AI Launches jina-reranker-v3 to Process 64 Documents

Jina AI has launched jina-reranker-v3, a multilingual listwise reranking model designed to evaluate multiple candidate documents simultaneously. Built on the 28-layer Qwen3-0.6B backbone, this 0.6-billion-parameter model introduces what the company calls a "last but not late interaction" architecture. Instead of running a traditional cross-encoder separately for every query-document pair, the new model packs a single query and up to 64 candidate documents into a shared 131,000-token context window, scoring them all in a single model pass.

The model achieves strong performance across several retrieval benchmarks. It scored 61.94 nDCG@10 on BEIR, outperforming competing rerankers that are 2.5 to 10 times larger. It also demonstrated high accuracy on other datasets, scoring 78.58 on HotpotQA and 94.01 on FEVER, alongside strong results on MIRACL, MKQA, and CoIR code retrieval. The model is available under a CC BY-NC 4.0 license and can be accessed via GGUF, MLX, and a hosted API. Additionally, Jina AI has already made its successor, version 3.5, available as a drop-in upgrade.

For search and retrieval practitioners, this listwise design reduces the repeated computations typically required by second-stage reranking. Because the model uses causal attention, later candidate documents can incorporate information from earlier ones in the sequence. However, this means document order matters, and developers should test shuffled candidate orders to ensure ranking stability. While the small 0.6-billion-parameter size lowers weight memory requirements, processing long candidate lists can still demand significant GPU memory. The model also offers optional 256-dimensional document embeddings conditioned on the query, though these are not suitable as static corpus embeddings.