BGE (BAAI General Embedding) from the Beijing Academy of Artificial Intelligence is the strongest family of open-source embedding models. BGE-M3 in particular supports multiple retrieval modes (dense, sparse, multi-vector) in a single model and handles 100+ languages. Self-hosted and free — the go-to choice for cost-sensitive or privacy-sensitive deployments.
BGE models are distributed on Hugging Face. You run them via sentence-transformers (Python) or Hugging Face Transformers. BGE-M3 produces three representations per text: dense vectors (for cosine similarity), sparse vectors (for BM25-like keyword match), and multi-vector (for late interaction). Self-hosting on a GPU is straightforward, and inference can be quantized for CPU deployment.
Free to self-host. You pay GPU costs: a single T4 can serve tens of millions of embeddings per day for BGE-small, or millions for BGE-large. Open source under permissive licenses.
Cost-sensitive RAG deployments, privacy-regulated industries, and research. Growing rapidly in 2026 as open embeddings close the gap with commercial options.