Researching local AI on open hardware.
We're a small collective building navi — an agent-native Linux distro where your AI never leaves your machine. Open weights, open systems, open hardware.
We're researching and building models for every machine: koe (声, “voice”, 270M–1B) for the weakest hardware, alice (1–3B) for personality, and agentic lain (8B+) for machines with headroom. Same program at every scale: open weights, perma-computing quants, validated on real hardware. Nobody's hardware is below AI. xxvoidsxx/koe-270m is the first.
Papers and deep writeups on local AI — quantization for weak hardware, fine-tuning tiny models, and everything we learn validating it on real machines. Published here alongside the models they describe. Read the papers.
Your hardware is the datacenter.