Jina's Reranker v2 Hits 15x Faster Search Across 26 Languages

Reranking is the last mile of a retrieval pipeline, the step where a fast vector search either becomes precise or falls apart. Jina Reranker v2 Base Multilingual is a compact cross-encoder built for that job, and it has quietly become one of the most downloaded reranking models on Hugging Face with over 1.2 million pulls.

The model takes a query and document pair as input and returns a relevance score. It weighs in at 278M parameters, uses BF16 tensors, supports a 1024-token context window, and was trained on query-document pairs across dozens of languages.

What it actually does

A bi-encoder embedding model scores query and document independently, which is fast but loses fidelity. A cross-encoder runs both through the transformer together, letting attention compare them token by token. That is more expensive, so you typically use it to rerank the top 50 to 200 hits from a vector search rather than scoring your whole corpus.

The v2 release pushes reranking well beyond plain text search. Compared with the previous jina-reranker-v1-base-en, v2 shows competitive numbers on benchmarks covering text retrieval, multilingual retrieval, function-calling-aware and text-to-SQL-aware reranking, and code retrieval. In practice, the same model can rerank passages, API tool calls for an agent, SQL query candidates, and code snippets.