Jina Reranker v3
Jina Reranker v3 introduced the company's last-but-not-late document ranking method. The 0.6-billion-parameter model jointly processes a query and candidate list rather than scoring only isolated query-document pairs.
How the model works
Its Qwen3 backbone uses causal self-attention across candidates. Last-token document representations feed a projection network and relevance calculation. Jina documents up to 64 documents within the approximately 131K-token context. Candidate interaction separates this design from ColBERT, which encodes documents independently before matching vectors.Jina Reranker v3 model details
Use and access
Full, GGUF and MLX weights remain published. Jina's current card recommends v3.5 for new projects and describes it as a compatible API upgrade. This entry retains v3 as an identifiable release for existing integrations and reproducibility. CC BY-NC 4.0 applies to downloaded weights; managed API or commercial licensing is a separate access route.Jina Reranker v3 model detailsJina Reranker v3 technical overview
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | API, Downloadable |
| Model identifiers | jina-reranker-v3jinaai/jina-reranker-v3jina-reranker-v3-GGUFjina-reranker-v3-mlx |
| Context window | 131,072 tokens 131K-class combined query and document-list context; limits also depend on candidate count. Jina Reranker v3 model details |
| Parameters | Approximately 0.6 billionJina Reranker v3 model details |
| Weight license | CC BY-NC 4.0Jina Reranker v3 model details |
| Base model | Qwen3-0.6BJina Reranker v3 model details |
Related models
- Jina Reranker v3.5
Jina AI · Jina Reranker
A listwise search ranker tuned for specialist documents and structured data.
Launch date API · Downloadable - Jina ColBERT v2
Jina AI · Jina ColBERT
A multilingual retriever that keeps separate vectors for document tokens.
Launch date API · Downloadable