jina-reranker-v2-base-multilingual

jina-reranker-v2-base-multilingual is a 278M-parameter cross-encoder reranker supporting 100+ languages with a 1,024-token context window (524,288 tokens with Flash Attention 2). It was the first Jina reranker to break the English-only limitation, delivering state-of-the-art performance on AirBench and strong results on structured data tasks including function calling and SQL schema matching. It processes documents 15× faster than comparable models.

The model employs a cross-encoder architecture enhanced with Flash Attention 2, enabling direct comparison between queries and documents across language barriers. The 278M-parameter BERT-based encoder uses symmetric bidirectional attention with ALiBi positional encodings for the 1,024-token base context (extendable to 524,288 tokens with Flash Attention 2). Training followed a four-stage progressive process: (1) English language capabilities, (2) cross-lingual alignment, (3) multilingual data incorporation, (4) hard-negative refinement. This staged approach prevented the catastrophic forgetting that often occurs when adding multilingual data to a single-language model. The Flash Attention 2 implementation is a key efficiency innovation, enabling 15× higher throughput than bge-reranker-v2-m3.

The model achieves state-of-the-art performance on the AirBench leaderboard for RAG systems and strong results on MKQA (26 languages). It excels in structured data tasks: high recall on ToolBench (function calling) and NSText2SQL (SQL schema matching). Most impressively, it processes documents 15× faster than comparable models like bge-reranker-v2-m3, making it practical for real-time applications. The 100+ language support and structured-data capabilities make it uniquely suited for agentic RAG systems and multilingual API documentation search. In 2026, jina-reranker-v3.5 supersedes this model with listwise interaction, 131K context, and domain-specific training.

The model requires a CUDA-capable GPU and is accessible through the Jina Reranker API, major RAG frameworks (Haystack, LangChain), and cloud marketplaces (AWS, Azure, GCP). It excels in scenarios requiring precise understanding across language barriers and data types: global enterprises working with multilingual content, API documentation search, and code retrieval across multilingual codebases. With the 1,024-token context window and automatic chunking for longer documents, it handles extended content efficiently. Use this model when you need to enhance search accuracy across languages, require function-calling capabilities for agentic RAG, or want to improve code search across multilingual codebases. For new projects, jina-reranker-v3.5 offers listwise ranking, 131K context, and domain-specific improvements.