BGE-M3
BAAI's MIT-licensed embedding model for 100+ languages and 8,192-token inputs, with dense, sparse and multi-vector search in one model.
In short. BGE-M3 turns text in more than 100 languages into numbers that capture its meaning, so software can search by meaning instead of by exact words. One model gives three kinds of search signal at once, and you can combine them for better results. It dates from early 2024, yet it is still one of the most downloaded embedding models on Hugging Face, and its license is MIT.
Summary
The Beijing Academy of Artificial Intelligence (BAAI) released BGE-M3 on 30 January 2024 as a member of its BGE series, and described it in the paper "M3-Embedding". The three Ms stand for multi-linguality (more than 100 working languages), multi-granularity (inputs from short sentences to documents of up to 8,192 tokens) and multi-functionality: the same model can do dense retrieval, sparse (keyword-weight) retrieval and multi-vector retrieval. The weights are an XLM-RoBERTa-style encoder with 24 layers and a hidden size of 1,024, stored as a 2.27 GB FP32 checkpoint, which implies roughly 568M parameters (our estimate from the file size). The paper says it was trained with self-knowledge distillation, where the relevance scores of the three retrieval modes together act as the teacher signal.
The model card recommends hybrid retrieval followed by re-ranking. The sparse output gives BM25-like token weights at no extra cost when you compute the dense vectors, so you can mix both in one search, and a cross-encoder such as BAAI's bge-reranker models can then re-score the top results. The card links hybrid-search examples for Vespa and Milvus, and the code lives in BAAI's FlagEmbedding repository on GitHub. The license is MIT.
BGE-M3 is now almost three years old, so check a current leaderboard rather than assuming it is the best choice. It stays popular (Hugging Face showed about 29 million downloads in the 30 days to 11 November 2026) because it is small, permissively licensed, multilingual and flexible, and because most vector-search tools already support it. Its 8,192-token input limit is shorter than the 32K of newer models such as Qwen3-Embedding.
Good for
- Multilingual semantic search and retrieval-augmented generation across 100+ languages
- Hybrid search that mixes dense, sparse and multi-vector scores from a single model
- Long passages and documents of up to 8,192 tokens
- A small, MIT-licensed default that runs on modest hardware
Watch out for
- Dates from January 2024, and newer embedding models may score higher, so test on your own data
- Vectors are not interchangeable between models, so switching means re-indexing
- The card recommends adding a re-ranker after retrieval for the best accuracy
- The 8,192-token limit is shorter than the 32K of newer models
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of BGE-M3. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| BGE-M3 FP32 | 568M | 2.27 GB | 2.61 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
- BGE-M3 vs Granite Embedding R2: which embedding model should you use?
- Best open-source embedding models in 2026
Frequently asked questions
- What is BGE-M3?
- BGE-M3 turns text in more than 100 languages into numbers that capture its meaning, so software can search by meaning instead of by exact words. One model gives three kinds of search signal at once, and you can combine them for better results. It dates from early 2024, yet it is still one of the most downloaded embedding models on Hugging Face, and its license is MIT.
- Can I use BGE-M3 commercially?
- Yes. BGE-M3 is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is BGE-M3 open source?
- BGE-M3 is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run BGE-M3 on my own computer?
- Yes. The smallest variant, BGE-M3, needs about 2.61 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was BGE-M3 released?
- BGE-M3 was released by BAAI on January 30, 2024.
- How long a context window does BGE-M3 have?
- BGE-M3 supports a context window of up to 8K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- BGE-M3 model card (Hugging Face) huggingface.co
- FlagEmbedding repository, release news (GitHub) github.com
- M3-Embedding paper (arXiv 2402.03216) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3-Embedding Alibaba Qwen | 600M – 7.57B | 32K | Apache 2.0 (Permissive) | 1.38 GB Laptop or phone | |
| Harrier OSS v1 Microsoft | 270M – 27B | 32K | MIT (Permissive) | 0.621 GB Laptop or phone | |
| Granite Embedding R2 IBM | 48M – 312M | 32K | Apache 2.0 (Permissive) | 0.115 GB Laptop or phone | |
| Jina Embeddings v5 Jina AI | 239M – 677M | 32K | CC BY-NC (Non-commercial) | 0.552 GB Laptop or phone |