BGE Reranker v2

BAAI/bge-reranker-v2-m3 is a multilingual cross-encoder reranking model from the Beijing Academy of Artificial Intelligence, part of the BGE (BAAI General Embedding) family. Unlike bi-encoder embedders that compare precomputed vectors, it reads the query and each candidate passage together with cross-attention and emits a direct relevance score — the standard second-stage upgrade for RAG pipelines. At 568M parameters on XLM-RoBERTa-large with 8192-token input, it covers 100+ languages and is widely pinned as the default quality reranker in open-source retrieval stacks.

The honest picture: the weights are about 2.3GB, so it is a real deployment cost next to 100MB-class MiniLM rerankers, and CPU inference runs seconds per query batch rather than milliseconds. Apache 2.0 licensing keeps it commercially usable. For teams that can afford the latency and memory, measured Hit@1 and MRR gains over smaller rerankers are substantial — but it is a quality-for-compute trade, not a free lunch.