BGE M3

BGE M3 is an embedding model developed by BAAI, positioned as a unified solution for multilingual and multi-granularity retrieval. Its three "M" design pillars combine Multi-Linguality, spanning more than 100 languages, Multi-Functionality, uniting dense, multi-vector, and sparse retrieval in a single model, and Multi-Granularity, accommodating inputs from short queries to long documents. This combination makes it suited for hybrid search pipelines where teams want one backbone rather than separate encoders for lexical and semantic matching.

As an open-weight release under the MIT License, BGE M3 can be redistributed, fine-tuned, and integrated into production retrieval stacks with permissive terms. The model is designed to achieve strong retrieval performance across diverse benchmarks, giving practitioners a flexible foundation for semantic search, reranking augmentation, and cross-lingual information retrieval tasks. For teams building hybrid retrieval systems that need broad language coverage without maintaining multiple specialized encoders, BGE M3 offers a consolidated approach backed by permissive licensing.