FlagEmbedding (BGE)

Behind many production RAG stacks sits FlagEmbedding, the retrieval toolkit from the Beijing Academy of Artificial Intelligence (BAAI) that develops the BGE model family. It spans the original BGE dense embedders in small, base, and large sizes for English and Chinese, the widely deployed BGE-M3 model that combines dense, sparse, and multi-vector retrieval across more than 100 languages with inputs up to 8,192 tokens, the bge-reranker-v2 cross-encoders, and the BGE-VL multimodal series released in March 2025 for visual search. BGE models have topped the MTEB and C-MTEB leaderboards and rank among the most downloaded embedding models on Hugging Face. The library installs with pip install -U FlagEmbedding and covers inference, evaluation, and fine-tuning, with an optional finetune extra pulling in training dependencies; smaller variants run acceptably on CPU while fine-tuning wants a GPU. Everything is MIT licensed, and the project, at around 12,000 GitHub stars, documents the models at bge-model.com.