bge-reranker-v2-m3

bge-reranker-v2-m3

Scores how well a passage answers a query so you can reorder search results, across many languages. Takes a question and a document and returns one relevance score. For people building search and retrieval pipelines.

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-reranker-v2-m3 is a multilingual reranker from BAAI, built on the bge-m3 model. It does not turn texts into vectors. It reads a query and a passage together and returns a single score, where a higher number means a closer match. A sigmoid step can squeeze that score into a range from 0 to 1.

Who it suits: It suits developers who already fetch candidate passages for a query and want a smarter final ordering.

Industry | Who uses it | How they use it | Example
Search products | Search engineers | Reorder retrieved results by how well each one fits the query | Putting the best-matching article first
Question answering | ML engineers | Pick the passage that best answers a question | Choosing between candidate answers
Internal knowledge bases | Platform teams | Rank the documents an index returns | Sorting policy documents for one query
Multilingual sites | Content teams | Compare queries and passages across languages | Matching a question with text written in another language

You can skip it if you need embeddings, since this model gives pair scores and no vectors.

Where to use it:

• Python projects that use the FlagEmbedding package
• Search and retrieval pipelines that already produce a list of candidates
• Windows, macOS or Linux with Python

Use case | Who | What they get back
Rerank search hits | Search teams | A relevance score for every query and passage pair
Score one pair | Developers | A raw score, or a 0 to 1 value with normalizing turned on
Compare many pairs | Data teams | A list of scores for a batch of pairs

How to set it up: You need Python and the FlagEmbedding package. The model files themselves are optional to fetch by hand.

• Python on Windows, macOS or Linux
• The FlagEmbedding package
• About 2.3 GB for the F32 weights, or less with fp16 loading

1. Install FlagEmbedding.
2. Optionally download the model files with the Hugging Face tool.
3. Load the reranker in Python and score your pairs.

pip install -U FlagEmbedding
pip install -U huggingface_hub
hf download BAAI/bge-reranker-v2-m3

How to use it: Load the reranker, then pass it a query and a passage, or a list of such pairs. Turn on normalizing if you want a value between 0 and 1.

from FlagEmbedding import FlagReranker
reranker = FlagReranker('BAAI/bge-reranker-v2-m3', use_fp16=True) # Setting use_fp16 to True speeds up computation with a slight performance degradation

score = reranker.compute_score(['query', 'passage'])
print(score) # -5.65234375

# You can map the scores into 0-1 by set "normalize=True", which will apply sigmoid function to the score
score = reranker.compute_score(['query', 'passage'], normalize=True)
print(score) # 0.003497010252573502

Tool | What it does
FlagReranker | Loads the reranker model
compute_score | Scores one pair or a list of pairs
normalize=True | Maps each score into the range 0 to 1
use_fp16=True | Runs faster, with a slight loss of quality

Example: You score the question "what is panda?" against the word hi and against a sentence about the giant panda. The card shows about -8.19 for hi and about 5.26 for the panda sentence, so the second one ranks higher.

For newcomers and experts: If you are new to search, picture a search engine that returns ten pages. The reranker reads each page next to the question and gives it a score, and you sort by that score. A first small try is the two-line score example above.

For people who know the field, this is a cross-encoder with 568M parameters, so each candidate needs its own forward pass and cost grows with the candidate list. The family also has gemma-based and MiniCPM layerwise versions for better quality or for choosing how many layers to run.

Title: Weight size of bge-reranker-v2-m3
F32: 2.3 GB
4-bit estimate: 284 MB

Your search returns candidate passages
The reranker reads the query with each passage
It returns one relevance score per pair
You sort by score and keep the best

Good to know: Other models in the same family exist, and the card suggests this one for multilingual work and for light, fast inference.

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the queries and passages you pass in
• Not for: producing embeddings, since it returns similarity scores instead of vectors

On Hugging Face, 13 November 2026

• Task: text classification
• Library: sentence-transformers
• Parameters: 568M
• Weights: about 2.3 GB in F32; about 284 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 14,685,230
• Likes: 1,301
• License: apache-2.0
• Quantized versions: 73
• Fine-tunes: 137
• Tags: sentence-transformers, safetensors, xlm-roberta, text-classification, transformers, text-embeddings-inference, multilingual

Links

• Model: https://huggingface.co/BAAI/bge-reranker-v2-m3
• Files: https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download BAAI/bge-reranker-v2-m3

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