We're a two-person lab: Thomas van Dongen and Stéphan Tulkens. We build small open-source models and the tools around them, so that embedding, search and data preprocessing run on any CPU, without a GPU or an API.
We think that if you make models fast enough, you unlock new possibilities. A model that answers in milliseconds on a laptop becomes a function you run: inside an agent loop, over every row of a dataset, or on a machine with no network. Most of what people use embeddings for doesn't need a large model. It needs a fast one that is good enough.
- model2vec: distill a sentence transformer into a static embedding model that is up to 500x faster.
- potion: our state-of-the-art static embedding models.
- semble: fast, accurate code search for agents.
- semhash: semantic data deduplication and filtering.
- vicinity: a unified interface for many nearest-neighbor search backends.
- model2vec-rs: (even) faster model2vec inference in Rust.
You can find us on: minish.ai · Hugging Face · Discord · LinkedIn
