BGE-M3
Partner VerifiedPublisher: BAAI · Category: Embedding
Parameters
560M
Context Window
8K
Default Size
1.2 GB
Downloads
390,200
Multi-lingual, multi-functionality embedding model supporting dense, sparse, and multi-vector retrieval.
PULL MODEL
eq pull bge-m3
RUN INFERENCE
eq run bge-m3 "Compute similarity score between query and context"
Attention Architecture
Utilizes Grouped-Query Attention (GQA) with 16-token fixed Paged KV Cache block allocation for zero-fragmentation memory residency.
ONNX & EQC Compatible
Compiles directly through the EQC toolchain into standalone .eqx binary packages with fused SwiGLU kernels.
Multi-Hardware Support
Auto-detects CUDA RTX/A100/H100, Apple Silicon Metal Performance Shaders, or AVX-512 CPU execution backends.