MinishLab/model2vec

Distills sentence transformers into compact static embeddings through lightweight PCA-based dimensionality reduction, enabling 500x faster inference with minimal performance loss. Supports distillation from any Sentence Transformer, token-level embeddings for sequence tasks, and fine-tuning classification heads directly on the static representations. Integrates with HuggingFace Hub for pre-trained models and supports BPE/Unigram tokenizers with optional int8 quantization for further size reduction.

2,008 stars and 582,040 monthly downloads. Used by 7 other packages. Actively maintained with 7 commits in the last 30 days. Available on PyPI.