BGE-M3 stands out for its Multi-Functionality (simultaneous dense, sparse, and multi-vector retrieval), Multi-Linguality (100+ languages), and Multi-Granularity (up to 8,192-token documents). It enhances retrieval pipelines by enabling hybrid retrieval (e.g., combining dense embeddings with BM25-like sparse weights) and re-ranking for higher accuracy. The model integrates seamlessly with tools like Vespa and Milvus, and its unified fine-tuning supports diverse retrieval methods. Recent updates include improved MIRACL benchmark performance and multilingual long-document datasets (MLDR).
