Model2Vec

Model2Vec is an open-source library from the Minish Lab team of Stephan Tulkens and Thomas van Dongen that turns a sentence transformer into a small static embedding model. It precomputes one fixed vector per token plus light post-processing, then averages token vectors to embed a sentence, which shrinks the model by up to fifty times and speeds inference by hundreds of times with a modest quality drop. The result suits classification, search, and on-device work where a full transformer is too slow or heavy.

Key Capabilities:

• Distillation that converts any sentence transformer into a static model

• Static token embeddings that run without a neural forward pass at inference

• Model size reductions up to fifty times the original transformer

• Inference speedups of hundreds of times on CPU

• Pretrained potion models, including a multilingual variant across many languages

• Training support for fine-tuning lightweight classification models