bge-m3

Multilingual BGE embedding model that supports dense retrieval, lexical retrieval, and multi-vector retrieval in one checkpoint.

BAAI · BGE

Editorial review

Reviewed byOpenSourcesAI EditorialLast updatedAugust 2026SourcesExact model card

Model checkpoints, context windows, provider support, local runtime compatibility, and license terms can change quickly. Verify the exact model card before production or commercial use.

Best for

RAG teams that want one embedding model for multilingual search, long documents, and retrieval experiments beyond plain dense vectors.

Who should use it

RAG teams that want one embedding model for multilingual search, long documents, and retrieval experiments beyond plain dense vectors.
Builders who want local or self-hosted testing options.
RAG builders comparing retrieval quality, latency, and multilingual coverage.

Common workflows

Multilingual embedding and retrieval workflows
embedding workflows
rag workflows
retrieval workflows
semantic search workflows

Deployment and hardware notes

~2.27 GB in fp32 (568M parameters), ~1.14 GB in fp16 — fine on any modern GPU, workable on CPU for small collections.

License and usage notes

MIT. Open weights. Verify the exact model card and license terms for the checkpoint or hosted provider you use.

Strengths

Open weights model option for BGE workflows.
RAG teams that want one embedding model for multilingual search, long documents, and retrieval experiments beyond plain dense vectors.
No query instruction needed — unlike the BGE v1.5 line, queries and passages are embedded bare, which simplifies pipelines.

Limitations

The 8,192-token window and three retrieval modes cost speed: it is several times slower per document than the 512-token BGE embedders, and the multi-vector output multiplies index storage if you use it.
~2.27 GB in fp32 (568M parameters), ~1.14 GB in fp16 — fine on any modern GPU, workable on CPU for small collections.
Context window and limits: 8,192 tokens.
Verify the exact model card, provider docs, license, and serving support before production use.

Local workflow notes

No query instruction needed — unlike the BGE v1.5 line, queries and passages are embedded bare, which simplifies pipelines.

Local runtimes: Ollama where supported, LM Studio where supported, llama.cpp where supported, Transformers

Platforms: Windows, macOS, Linux

Embedding spec

Memory~2.27 GB in fp32 (568M parameters) · ~1.14 GB in fp16Output1024-dimension vectorsMax input8,192 tokens

One checkpoint, three retrieval signals: 1024-dimension dense vectors, sparse lexical weights, and ColBERT-style multi-vectors, across 100+ languages and inputs up to 8,192 tokens. The default pick when documents outrun 512 tokens or the corpus is multilingual.

Sources to verify

Exact model card

Related resources

Continue with model source notes, local tools, and implementation guides related to this model.

BGE family hub Embedding models shortlist Compare BGE vs E5 RAG for documents stack AI tools directory

Hardware~2.27 GB in fp32 (568M parameters)RuntimeOllama or LM Studio where supported, llama.cpp, Transformers, vLLMContext8,192 tokensLast updated2026

Exact model card →