BGE-M3
BAAI · Embedding · 568M · 8k context · Released 30 January 2024
Commercial use permitted
Open weights
Runs on CPU
Apple Silicon
A multilingual embedding model that is unusual for doing dense, sparse, and multi-vector retrieval from a single model, across more than 100 languages, with an 8192-token input. MIT licensed and a common default for local RAG.
Strengths
- Dense, sparse, and multi-vector retrieval in one model
- Over 100 languages, useful for multilingual document sets
- MIT licensed, and a long 8192-token input
Weaknesses
- Heavier and slower than tiny embedders like nomic-embed
- Multi-vector retrieval adds storage and retrieval complexity
- A general model, so a domain-tuned embedder may beat it on narrow tasks
Hardware requirements
| Quantisation | Approx. VRAM | Notes |
|---|---|---|
| FP16 | ~1.2GB | Full precision, runs comfortably on any GPU or CPU |
| Int8 | ~0.7GB | Lighter, minimal quality loss for retrieval |
Also runs on CPU (slower). Optimised builds available for Apple Silicon.
Licence
MIT — read the licence
Availability
- Official page
- Hugging Face
- ollama run bge-m3
Recommended for
- Multilingual retrieval for RAG systems
- Document sets that mix languages
- A capable default when you are not sure which embedder to use
Run it with
Catalogue entry last verified 30 July 2026. Specifications change; verify anything you are about to spend money on.