bge-m3

bge-m3

Turns text in more than 100 languages into vectors for search, from short sentences to documents of 8192 tokens. It supports dense, multi-vector and sparse retrieval in one model. By BAAI.

Made by BAAI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: BGE-M3 is an embedding model from BAAI that converts text into vectors so you can search and compare passages. Its three traits are many functions, many languages and many input sizes.

It offers three retrieval modes from one model: dense, multi-vector and sparse. It supports over 100 languages, and the text you give it can be a brief sentence or a long document that runs to 8192 tokens.

Who it suits: Developers who build search, question answering or retrieval-augmented generation over text in many languages, and researchers who compare retrieval methods.

Industry | Who uses it | How they use it | Example
Software | Teams building RAG systems | Embed documents and queries for retrieval | Find relevant passages across languages
Enterprise search | Knowledge base builders | Match questions to long documents | Search files of up to 8192 tokens
Research | Retrieval researchers | Test dense, sparse and multi-vector modes | Compare hybrid retrieval against a single mode

Skip it if you need a model that writes text, since this one only produces vectors and token weights for matching.

Where to use it:

• Python programs that use sentence-transformers or the FlagEmbedding project
• Vector databases such as Milvus and Vespa, which the card says support its hybrid retrieval
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Semantic search | Search developers | Dense vectors that match meaning across languages
Keyword-style matching | Teams using hybrid retrieval | Sparse token weights similar to BM25
Long document retrieval | Knowledge base teams | Embeddings for inputs of up to 8192 tokens

How to set it up:

• Python and pip
• The sentence-transformers or FlagEmbedding library to run the model
• A vector database if you want to store and search the vectors

1. Install the Hugging Face tools.
2. Download the model files.
3. Follow the model card and the FlagEmbedding project for the code that creates embeddings.

pip install -U huggingface_hub
hf download BAAI/bge-m3

How to use it: Embed your documents and your queries with the model, then compare them. The card recommends a pipeline of hybrid retrieval followed by re-ranking.

Mode | What it does
Dense retrieval | Compares whole-text vectors for meaning
Multi-vector retrieval | Keeps several vectors per text for finer matching
Sparse retrieval | Gives token weights, similar to BM25, at no extra cost when you compute dense embeddings
Re-ranking | Uses a reranker model after retrieval to filter the selected text

Example: You embed a set of documents in several languages, then embed a question and rank the documents by similarity. A re-ranker such as bge-reranker then filters the top results for better accuracy.

For newcomers and experts: If you are new to embeddings, think of this model as a way to turn each piece of text into a list of numbers where similar meanings sit close together. Start by embedding two sentences and comparing them.

If you know the field, one checkpoint covers dense, multi-vector and sparse retrieval, so you can mix embedding and lexical signals without a second model. The card recommends pairing hybrid retrieval with a cross-encoder re-ranker.

Embed documents with BGE-M3
Retrieve with dense and sparse signals
Merge results in hybrid retrieval
Re-rank with a cross-encoder

Good to know:

• Hugging Face lists 313 quantized and 597 fine-tuned versions made by other people
• The card also lists related models, including bge-m3-unsupervised, bge-m3-retromae and the English-only bge-large-en-v1.5

Title: Maximum input length in tokens
bge-m3: 8192 tokens
bge-m3-unsupervised: 8192 tokens
bge-large-en-v1.5: 512 tokens

• License: MIT
• Cost to run: none named
• Data it reaches: the text you embed, plus the files you download
• Not for: generating text, since it produces vectors and token weights

On Hugging Face, 13 November 2026

• Task: sentence similarity
• Library: sentence-transformers
• Downloads in the last 30 days: 29,282,167
• Likes: 3,989
• License: mit
• Quantized versions: 314
• Fine-tunes: 597
• Tags: sentence-transformers, pytorch, onnx, xlm-roberta, feature-extraction, sentence-similarity, text-embeddings-inference

Links

• Model: https://huggingface.co/BAAI/bge-m3
• Files: https://huggingface.co/BAAI/bge-m3/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download BAAI/bge-m3

Then follow the usage in its model card. Nothing in this kit runs by itself.