The local intelligence lab

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.

Projects

• 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 · GitHub · Discord · LinkedIn