Michael Feil's bge-small-en-v1.5 Fork Pulls 2.3M Downloads Despite 33M Parameters
Michael Feil's ONNX-ready mirror of BAAI's bge-small-en-v1.5, the default model for the Infinity serving engine, crossed 2.3M monthly downloads.
Takeaways
- michaelfeil/bge-small-en-v1.5 hit 2.3M monthly downloads on Hugging Face.
- It is a deployment-ready mirror of BAAI's 33.4M-parameter English embedder, not a new model.
- Ships PyTorch, Safetensors, and ONNX weights in one repo for GPU or CPU serving.
- Acts as the default model for the Infinity embeddings serving engine.
- One-line CLI exposes an OpenAI-compatible embeddings API on port 7997.
- Best for English RAG, semantic search, clustering; capped at 512 tokens, English-only.
A community-maintained fork of one of the most popular open embedding models keeps quietly racking up traffic on Hugging Face. Michael Feil's repackaged version of BAAI's tiny English embedder, michaelfeil/bge-small-en-v1.5, pulled 2,312,577 downloads in the last month, putting it in the same league as many billion-parameter LLMs despite weighing in at just 33.4M parameters.
The repository mirrors BAAI/bge-small-en-v1.5 in a deployment-friendly layout, shipping PyTorch, Safetensors, and ONNX weights side by side so you can pull one artifact and serve it on either GPU or CPU without a separate conversion step. It exists because it is the stable default model for the Infinity project, Feil's open-source embeddings server.
Why a repackaging job draws millions of pulls
BAAI's original BGE weights are excellent, but they were published as a PyTorch checkpoint aimed at research use. Feil's repo adds the plumbing needed for production: ONNX graphs, an optimum-compatible layout, and a model card that documents how to run it inside Infinity, a high-throughput serving engine for embeddings, rerankers, CLIP, CLAP, and ColPali.