bge-m3

This repository contains a BF16 build of BAAI/bge-m3, together with a Furiosa Executable Bundle (FXB) for running it on FuriosaAI RNGD with Furiosa-LLM. The base model also runs on other frameworks (such as FlagEmbedding, Sentence Transformers, and Transformers); for usage with those, see the upstream BAAI/bge-m3 model card.

Overview

BGE-M3 is a multilingual text-embedding model developed by BAAI on the XLM-RoBERTa encoder architecture. The upstream model unifies dense, sparse lexical, and multi-vector retrieval. Furiosa-LLM exposes its 1,024-dimensional, CLS-pooled dense embeddings for semantic search, retrieval, and similarity matching. Its intended use is the same as the upstream BAAI/bge-m3, and it is released under the MIT License.

• Architecture: XLM-RoBERTa (dense encoder), XLMRobertaModel
• Input / Output: Text / Dense embeddings (vector)
• Supported Inference Engine: Furiosa LLM
• Supported Hardware: FuriosaAI RNGD

Quantization

The FuriosaAI build uses BF16 weights without a lower-bit quantization scheme.

Parallelism Strategy

On RNGD, bge-m3 runs with a tensor-parallel size of 8 PEs, which maps to a single RNGD card (8 PEs per card).

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-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 an OpenAI-compatible /v1/embeddings endpoint. BGE-M3 does not require an instruction prefix for queries. Request dense embeddings with curl:

curl http://localhost:8000/v1/embeddings \
    -H "Content-Type: application/json" \
    -d '{
    "model": "furiosa-ai/bge-m3",
    "input": [
      "What is BGE-M3?",
      "BGE-M3 is a multilingual text-embedding model."
    ]
    }' \
    | python -m json.tool

Because the endpoint is OpenAI-compatible, you can also use the OpenAI Python client:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

response = client.embeddings.create(
    model="furiosa-ai/bge-m3",
    input=[
        "What is BGE-M3?",
        "BGE-M3 is a multilingual text-embedding model.",
    ],
)

for data in response.data:
    print(f"Index {data.index}: {len(data.embedding)} dimensions")

Advanced Usage

For offline use, load the model with the LLM constructor (the FXB shipped in the repo is discovered automatically) and call embed to obtain L2-normalized dense vectors. Their dot product is therefore the cosine similarity:

from furiosa_llm import LLM

texts = [
    "What is BGE-M3?",
    "BGE-M3 is a multilingual text-embedding model.",
]

with LLM("furiosa-ai/bge-m3") as llm:
    outputs = llm.embed(texts)
    embeddings = [output.outputs.embedding for output in outputs]

similarity = sum(a * b for a, b in zip(*embeddings, strict=True))
print(f"Cosine similarity: {similarity:.4f}")

Learn more

• Furiosa-LLM Server (furiosa-llm serve) — full OpenAI-compatible API reference, including the Embeddings API
• Furiosa-LLM — Furiosa-LLM documentation and API reference
• BAAI/bge-m3 — upstream model card