BGE-M3 is built by the Beijing Academy of Artificial Intelligence (BAAI) as a next-generation text embedding model that breaks from the tradition of single-purpose retrievers. Where most embedding models specialize in one retrieval strategy, BGE-M3 unifies three distinct approaches—dense retrieval, multi-vector retrieval, and sparse retrieval—running them simultaneously within a single forward pass. This design enables the model to generate dense embeddings, weighted token-level sparse vectors, and multi-vector representations without adding extra computational steps. The architecture is further refined through self-knowledge distillation, a training technique that lets the model transfer its own learned signal to improve adaptability and accuracy across datasets. With 569 million parameters organized into 1024-dimensional output vectors, BGE-M3 handles everything from brief queries to full documents in a single pass.

The model is pretrained on text spanning over 170 different languages, making it a strong fit for global retrieval pipelines where monolingual models would fall short. Its design emphasizes practical versatility: developers can apply it to hybrid search pipelines combining embedding similarity with lexical signals like BM25, or layer on a cross-encoder reranker to refine top results. BGE-M3 achieved state-of-the-art scores on multilingual benchmarks including MKQA and MLDR, validating its strength in cross-lingual retrieval tasks. Being released as an open-weight model under the MIT license means teams can host it directly, fine-tune it on domain-specific corpora, or integrate it with vector databases such as Milvus for production-scale similarity search.