BAAI/bge-reranker-v2-m3

A powerful, multilingual cross-encoder model designed to optimize search relevance by directly scoring query-passage pairs.

Developed by the Beijing Academy of Artificial Intelligence (BAAI), bge-reranker-v2-m3 is a state-of-the-art cross-encoder model built to refine retrieval-augmented generation (RAG) pipelines. Unlike standard bi-encoder embedding models that process queries and documents separately, this model ingests them simultaneously to output a precise, direct relevance score. It excels in multilingual environments, supports a massive 8,192-token context window, and is highly optimized for deployment across diverse environments (including ONNX and Ollama runtimes).

https://huggingface.co/BAAI/bge-reranker-v2-m3