Model2Vec

Distillation in about thirty seconds on a CPU is the headline trick of Model2Vec from MinishLab: it converts any Sentence Transformer into a static embedding model by passing a vocabulary through the teacher once and storing the resulting token vectors, with PCA and frequency weighting applied on top. The output needs no neural network at inference, which makes distilled models up to 500x faster and up to 50x smaller, from the 1.8M-parameter potion-base-2M up to the 128M-parameter potion-multilingual-128M released in May 2025 covering 101 languages, at a modest accuracy cost on MTEB. No training data is required for basic distillation, and optional extras add fine-tuning and training pipelines. Installation is pip install model2vec, inference runs anywhere Python does, and the models plug into Sentence Transformers, LangChain, and the Hugging Face Hub. Code and released models are MIT licensed. Teams use it to cut embedding costs in latency-sensitive or high-volume pipelines where a full transformer is overkill; the repository has about 2,200 stars.