Jina Reranker

Reranking Model (Open Source + API) J AI Processing & RAG

Basic Information

Product Description

Jina Reranker is a series of reranking models launched by Jina AI, continuously iterating and upgrading from v1 to v3. The latest jina-reranker-v3 is a 0.6B parameter multilingual document reranking model, featuring an innovative "last but not late interaction" architecture, achieving the highest score of 61.94 nDCG-10 on the BEIR benchmark. Jina also introduced jina-reranker-m0, the world's first multimodal multilingual reranking model, capable of handling visually rich document images.

Core Features/Highlights

Jina Reranker V3

  • Innovative Architecture: "Last but not late interaction"—causal self-attention for queries and documents within the same context window
  • Efficient Ranking: Single inference can handle up to 64 documents simultaneously
  • Ultra-Long Context: 131K token context window
  • Highest BEIR Score: 61.94 nDCG-10, surpassing all evaluated rerankers
  • Lightweight: Based on Qwen3-0.6B, only 28 layers Transformer + MLP projector (1024→512→256)
  • 4.88% Improvement Over V2: Significant precision enhancement

Jina Reranker V2

  • Agentic RAG Design: Specifically built for agentic RAG
  • Function Calling: Supports Function-calling
  • Multilingual: Supports 100+ languages
  • Code Search: Specialized code search capability
  • 6x Speedup: 6 times faster than v1
  • 15x Throughput: 15 times higher document throughput than bge-reranker-v2-m3
  • Flash Attention: Built-in flash attention mechanism
  • 1024 Token Context

Jina Reranker M0 (Multimodal)

  • Multimodal Reranking: The world's first multimodal multilingual reranking model
  • Visual Documents: Capable of handling document images containing text, charts, tables, and various layouts
  • 29 Languages: Multilingual visual document support

Business Model

  • Model Weights: Partially open-sourced on Hugging Face
  • API Service: Paid API available through jina.ai
  • MLX Training: Supports training embedding and reranking models on Apple Silicon
  • Elasticsearch Integration: Official Elasticsearch integration of Jina reranking models in 2026

Target Users

  • RAG system developers
  • AI application teams requiring high-precision retrieval
  • Multimodal document retrieval scenarios
  • Code search application developers
  • Agent systems requiring Agentic RAG capabilities

Competitive Advantages

  • V3 achieves the highest score among all rerankers on the BEIR benchmark (61.94)
  • Innovative architecture design (different from traditional ColBERT's separate encoding)
  • V2 specifically designed for Agentic RAG, supporting function calling
  • M0 pioneers multimodal reranking
  • Capable of handling 64 documents simultaneously, highly efficient
  • Apple Silicon training support (MLX)
  • Official Elasticsearch integration

Limitations

  • Commercial license restrictions (CC-BY-NC-4.0 for some models)
  • API pricing less transparent than Cohere
  • Community size smaller than Cohere
  • V3 based on Qwen3 model, may have related license constraints

Relationship with OpenClaw Ecosystem

Jina Reranker provides multidimensional reranking capabilities for OpenClaw. V3's highest BEIR precision is suitable for scenarios requiring strict retrieval quality; V2's function calling and Agentic RAG design directly align with OpenClaw's agentic architecture; M0's multimodal reranking enables OpenClaw to retrieve and rank visual documents. MLX support also allows Mac users to locally train and run Jina reranking models.

External References

Learn more from these authoritative sources: