maxiboch/bge-m3-onnx

BGE-M3 ONNX (Full Multi-Vector)

ONNX conversion of BAAI/bge-m3 with full multi-vector support.

Converted using yuniko-software/bge-m3-onnx method.

Why This Conversion?

Most BGE-M3 ONNX conversions only output dense embeddings. This conversion preserves all three retrieval methods:

Output	Use Case
Dense vectors (1024-dim)	Semantic similarity search
Sparse vectors	Lexical/keyword matching (hybrid search)
ColBERT vectors	Late interaction retrieval

Multilingual Support

100+ languages including English, Chinese, Japanese, Korean, German, French, Spanish, Arabic, Hindi, and many more.

Model Info

Property	Value
Embedding Dimension	1024
Max Sequence Length	8192
Languages	100+
Model Size	~2.1 GB

Files

File	Description
bge_m3_model.onnx	Main model graph
bge_m3_model.onnx_data	External weights
bge_m3_tokenizer.onnx	ONNX tokenizer
tokenizer.json	HuggingFace tokenizer

Usage

import onnxruntime as ort
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("maxiboch/bge-m3-onnx")
session = ort.InferenceSession("bge_m3_model.onnx", 
    providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])

inputs = tokenizer("Your text here", return_tensors="np", padding=True, truncation=True)
outputs = session.run(None, dict(inputs))

dense_embeddings = outputs[0]   # Semantic embeddings
sparse_weights = outputs[1]     # For hybrid search
colbert_vecs = outputs[2]       # For late interaction

Credits

• Original model: BAAI/bge-m3
• Conversion method: yuniko-software/bge-m3-onnx
• Converted by: @maxiboch
• License: MIT