BGE Reranker v2 M3

BGE Reranker v2 M3#

BGE Reranker v2 M3 is a multilingual cross-encoder reranking model developed by BAAI on the XLM-RoBERTa encoder architecture. Given a query and a set of candidate documents, it produces relevance scores used to reorder retrieval results — a common second stage in retrieval-augmented generation (RAG) and search pipelines.

FuriosaAI publishes the upstream model under the furiosa-ai organization on the Hugging Face Hub, shipping a Furiosa Executable Bundle (FXB) for running it on FuriosaAI RNGD with Furiosa-LLM. The same upstream weights also run on other frameworks (such as FlagEmbedding, Sentence Transformers, and Transformers); for usage with those, see the upstream model card linked below.

For the related BGE embedding model see BGE-M3; for another reranker family see Qwen3-Reranker.

Available Models#

Model

Quantization

RNGD cards

Notes

furiosa-ai/bge-reranker-v2-m3

None (FP32 weights; BF16 FXB)

1

Multilingual cross-encoder reranker

  • Architecture: XLM-RoBERTa (dense encoder), XLMRobertaForSequenceClassification

  • Task: Reranking

  • Input / Output: Text (query-document pairs) / Relevance score

  • Quantization: No lower-bit quantization is applied. The upstream checkpoint stores FP32 weights; Furiosa-LLM downcasts them to BF16 for W16A16 execution by the FXB.

Usage#

To run this model with Furiosa-LLM, follow the examples below after installing Furiosa-LLM and its prerequisites. You can use the model either online through the OpenAI-compatible server or offline through the Furiosa-LLM Python API.

Launch the server#

Serve the model by passing its furiosa-ai/<repo> identifier:

# Launch the server, listening on port 8000 by default
furiosa-llm serve furiosa-ai/bge-reranker-v2-m3

When the server is ready, you will see:

INFO:     Started server process [27507]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:8000 (Press CTRL+C to quit)

Basic Usage#

The server exposes a /v1/rerank endpoint compatible with the Cohere and Jina rerank APIs. Send a query and candidate documents with curl; the server returns the documents ordered by their sigmoid-normalized relevance_score:

curl http://localhost:8000/v1/rerank \
    -H "Content-Type: application/json" \
    -d '{
    "model": "furiosa-ai/bge-reranker-v2-m3",
    "query": "What is deep learning?",
    "documents": [
        "Deep learning is a subset of machine learning using neural networks.",
        "Python is a popular programming language for data science.",
        "Neural networks are inspired by biological neural networks."
    ]
    }' \
    | python -m json.tool

You can pass top_n to keep only the most relevant documents. To score query-document pairs directly instead of reranking, the server also exposes a /v1/score endpoint.

Advanced Usage#

For offline use, load the model with the LLM constructor (the FXB shipped in the repo is discovered automatically) and call score with a query and the candidate documents to obtain relevance scores:

from furiosa_llm import LLM

query = "What is deep learning?"
documents = [
    "Deep learning is a subset of machine learning using neural networks.",
    "Python is a popular programming language for data science.",
]

with LLM("furiosa-ai/bge-reranker-v2-m3") as llm:
    outputs = llm.score(query, documents)
    for document, output in zip(documents, outputs, strict=True):
        print(f"score={output.outputs.score:.4f}  {document}")

Learn more#