BGE-M3 is an advanced text embedding model developed by BAAI that excels in multi-functionality, multi-linguality, and multi-granularity. It supports dense retrieval, sparse retrieval (lexical weighting), and multi-vector retrieval (ColBERT-style), making it ideal for hybrid systems in retrieval-augmented generation (RAG). The model handles over 100 languages and supports long-text inputs up to 8192 tokens, offering flexibility across short queries and full documents. BGE-M3 was trained using self-knowledge distillation and unified fine-tuning strategies to align its performance across all modes. It achieves state-of-the-art results in several multilingual and long-document retrieval benchmarks, surpassing models from OpenAI in certain tests. Designed to integrate with tools like Milvus and Vespa, BGE-M3 enables efficient hybrid retrieval pipelines and downstream scoring via re-ranking models.
Features
- Supports dense, sparse, and multi-vector retrieval in one model
- Multilingual coverage across 100+ languages
- Handles sequences up to 8192 tokens
- Fine-tuned for document, sentence, and passage embeddings
- Outperforms OpenAI models in key multilingual benchmarks
- Plug-and-play integration with Vespa and Milvus for hybrid search
- Built with self-distillation and efficient batching for long texts
- Open-sourced under the MIT license with full documentation and examples