Bge M3 Onnx
bge-m3-onnx is an ONNX implementation of the BGE-M3 multilingual embedding model and tokenizer, enabling native execution in C, Java, and Python. This solution generates all three specific embedding types supported by BGE-M3: dense, sparse, and ColBERT vectors. It is designed for developers requiring high-performance, multi-vector retrieval capabilities without external dependencies or internet connectivity. The package supports cross-platform deployment and includes CUDA GPU acceleration for reduced latency in local embedding generation. Key features include full control over the embedding pipeline, offline operation, and verified cross-language compatibility ensuring consistent vector outputs across .NET, Java, and Python environments. The repository provides complete tooling for converting the original FlagEmbedding BAAI/bge-m3 model to ONNX format, including a Jupyter notebook guide, reference embedding generators, and test scripts to validate consistency across languages. Samples are provided for each su