Overview
BGE-M3, from the Beijing Academy of Artificial Intelligence, is named for three “M”s: multi-linguality, multi-functionality and multi-granularity. One model supports over 100 languages, inputs up to 8,192 tokens, and three retrieval methods — dense vectors, sparse lexical weights and multi-vector representations.
Highlights
- Support for over 100 languages
- Inputs up to 8,192 tokens
- Dense, sparse and multi-vector retrieval in one model
- MIT license
- Small enough to self-host on a single GPU
Typical uses
Self-hosted multilingual search and RAG systems where data must stay in-house.
Things to consider
- Self-hosting requires infrastructure for serving embeddings at scale.
- Using all three retrieval modes adds system complexity.
- Evaluate against newer embedding models on your data.
Specifications are taken from the provider’s official announcement, documentation or model card. Benchmark results depend on evaluation settings; see how AI benchmarks work.