FastEmbed

Embedding generation without the PyTorch baggage is the pitch behind FastEmbed, Qdrant's Python library that runs quantized transformer models through ONNX Runtime instead of a full deep learning stack. It covers dense text embeddings such as BAAI/bge-small-en-v1.5, sparse SPLADE++ vectors, ColBERT late interaction models, image embeddings, ColPali multimodal retrieval, and cross-encoder rerankers behind one consistent API. Because inference is CPU-first with a small dependency footprint, it suits serverless runtimes like AWS Lambda where installing PyTorch is impractical, and built-in data parallelism helps when encoding large datasets. Installation is pip install fastembed, with an optional fastembed-gpu package adding CUDA 12.x acceleration. The library integrates tightly with the Qdrant vector database client, which can call it automatically when upserting documents. Code is Apache-2.0 licensed, and the project, at around 3,100 GitHub stars, is the recommended embedding path for Qdrant users.