Haku

Haku is a small, independent lab that tunes small language models to run locally on Mac (on the GPU and Apple Neural Engine) and builds the software around them. Its approach centers on narrow task-specific models (dictation formatting, date resolution, tool-calling, page-control ranking), quantization to ternary / low-bit weights via quantization-aware training, and a frozen base plus small LoRA adapters. Inference runs through the company's own Metal engine and user data stays on the device. Products are sold as one-time purchases with lifetime updates, no subscriptions or accounts. Two apps (HakuDictate, Haku Terminal) ship on macOS 14+, with two more (Browse, Code) announced; Windows, iOS and Android versions are planned. Research results are published at lab.haku.sh.