jina-reranker-v3
Open Weights
Jina AI
Jina AI's 597M-parameter multilingual document reranker, built on Qwen3-0.6B with a 'last but not late' interaction architecture that scores up to 64 documents in a single forward pass. It reaches 61.94 nDCG@10 on BEIR while being roughly an order of magnitude smaller than generative listwise rerankers, and supports a 131K-token context.
Strengths
- 61.94 nDCG@10 on BEIR at only 597M parameters
- Scores up to 64 documents in a single forward pass via last-but-not-late interaction
- 131K-token context window
- Covers 93 languages, with 24 represented in the training set
- Roughly 10x smaller than generative listwise rerankers of comparable quality
- Open weights, so it can run on-premise
Caveats
- CC-BY-NC-4.0 - non-commercial use only without a separate licence
- A reranker, not a generative model - requires a first-stage retriever
- At 0.6B parameters it targets efficiency rather than absolute peak relevance
- Vendor
- Jina AI
- Context Window
- 131,072 tokens
- License
- CC-BY-NC-4.0
- Release Date
- 2025-10-01
- Modalities
- text
Capabilities
Vision
Audio
Video
Tool Use
Resources
Pricing
Open weights under CC-BY-NC-4.0 (non-commercial); also available through Jina's hosted Reranker API on a token-based plan.