bge-m3

Overview. BGE-M3 is a multilingual text-embedding model from the Beijing Academy of Artificial Intelligence (BAAI), released under the MIT licence. It turns text in more than 100 languages into embedding vectors for search and retrieval, with inputs up to 8192 tokens.

Key strengths. BGE-M3 gives strong multilingual retrieval quality from a compact 560M-parameter encoder and handles inputs up to 8192 tokens. Queries need no instruction prefix. Through Stav it returns dense embedding vectors on the standard embeddings API.

When to use it. Use BGE-M3 for multilingual semantic search, document retrieval and retrieval-augmented generation, especially when documents mix languages or run long. For RAG, pair it with a separate re-ranking stage, as BAAI recommends.

Limitations. BGE-M3 is an embedding encoder: it does not generate text, chat or call tools. Through Stav it returns dense vectors only; the sparse and multi-vector outputs described in the model card are not exposed. Inputs above 8192 tokens must be split.

Sovereignty note. BGE-M3 is an open-weight model, so it runs on the serving provider's own infrastructure and no traffic goes to its creator. Where a request is processed depends on the provider endpoint that serves it.