BGE Reranker v2 M3 is a cross-encoder reranking model that takes a query and passage together and directly outputs a relevance score rather than generating embeddings. Built on the multilingual BGE-M3 architecture, it applies cross-attention between the query and document to produce similarity signals that can be mapped to probabilities via sigmoid. This design makes it well-suited for improving search relevance by re-scoring retrieved passages in enterprise and multilingual search pipelines.

The model is developed by BAAI under the FlagEmbedding project and offers strong cross-language capabilities while maintaining fast inference speeds. With 568M parameters across 2.27 GB, it achieves competitive reranking performance while remaining lightweight enough for straightforward deployment. Its multilingual focus enables cross-language search scenarios, and its tight integration with the broader BGE embedding ecosystem makes it particularly effective for RAG pipelines requiring multilingual retrieval and ranking.