skip to content
gigarouter gigarouter
models / reranker · coming soon

Jina Reranker V3

jinaai/jina-reranker-v3

published Sep 2025 · updated Mar 2026

Jina Reranker V3 is a 0.6B parameter multilingual listwise document reranker that uses a novel last but not late interaction architecture to achieve state-of-the-art retrieval performance.

est. price
~$0.008
/ 1k docs · estimated, set at launch
API providers
0
downloads / mo
949.9K
license
cc-by-nc-4.0

specs

TaskDocument Reranking
ArchitectureLast but not late interaction (causal self-attention between query and documents)
Parameters0.6B total (0.44B non-embedding)
LicenseCC BY-NC 4.0

about this model

jinaai/jina-reranker-v3 is a 0.6B parameter multilingual listwise document reranker that uses a novel "last but not late interaction" architecture. Unlike late interaction models such as ColBERT which encode documents separately before multi-vector matching, this model applies causal self-attention between the query and all candidate documents within the same context window, extracting contextual embeddings from the last token of each document. This enables rich query-document interactions while keeping computational costs low.

Architecture diagram of jina-reranker-v3 showing query and documents processed jointly through transformer layers with a lightweight MLP projector.

Built on Qwen3-0.6B with 28 transformer layers and a lightweight MLP projector (1024→512→256 dimensions), the model processes up to 64 documents simultaneously within a 131K token context. Its 0.44B non-embedding parameters contribute to efficient inference while delivering state-of-the-art retrieval quality. On the BEIR benchmark, it achieves 61.94 nDCG@10 outperforming larger rerankers while being roughly 10× smaller than generative listwise models. The model maintains stable performance across different input document orderings, demonstrating robustness to permutation.

Multilingual capabilities are evaluated across 18 languages on MIRACL and 26 languages on MKQA, with scores of 66.83 and 67.92 respectively.

ModelSizeBEIRMIRACLMKQACoIR
jina-reranker-v30.6B61.9466.8367.9270.64
jina-reranker-v20.3B57.0663.6567.9056.14
jina-reranker-m02.4B58.9566.7568.1963.55
bge-reranker-v2-m30.6B56.5169.3267.8836.28
mxbai-rerank-base-v20.5B58.4055.3264.2465.71
mxbai-rerank-large-v21.5B61.4457.9467.0670.87
Qwen3-Reranker-0.6B0.6B56.2857.7065.3465.18
Qwen3-Reranker-4B4.0B61.1667.5267.5273.91
jina-code-embeddings-0.5b0.5B---73.94

This model is hosted as a managed, OpenAI-compatible API on gigarouter. For further technical details, see the arXiv paper.

best for

FAQ

What is the input format for the rerank API on gigarouter?

The API accepts a query string and a list of document strings via a POST request to the OpenAI-compatible endpoint, using your gigarouter API key.

How does this model compare in size and speed to other rerankers?

At 0.6B parameters, it is 10x smaller than generative listwise rerankers while achieving state-of-the-art BEIR performance (61.94 nDCG@10).

What is the license for Jina Reranker V3?

The model is licensed under CC BY-NC 4.0. Commercial use requires contacting Jina AI.

How many documents can the model process at once?

It can process up to 64 documents simultaneously within a 131K token context window.

What languages does the model support?

It supports multilingual retrieval, with strong performance on BEIR (English), MIRACL (18 languages), and MKQA (26 languages).

not yet live

We're benchmarking and onboarding Jina Reranker V3 as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.

related reranker models

compare all →