self

Run a local AI model server with a registry-driven workflow.

View the project on GitHub

Quick start

The preferred way to run self is through Docker. The development commands below are Docker-backed shorthand: self serve <model> means running the image's docker run --rm --gpus=all -p 8080:8080 ghcr.io/k0in/self:latest serve <model>.

docker run --rm -p 8080:8080 \
  --gpus=all \
  -v "$HOME/.ai-server/models:/models" \
  ghcr.io/k0in/self:latest serve kev:0.5b

Images are published to GHCR for Linux hosts:

Tag	Runtime
latest	NVIDIA CUDA 13
cuda-12	NVIDIA CUDA 12
cpu	CPU-only

CUDA images require an accessible NVIDIA GPU, a host NVIDIA driver, and the NVIDIA Container Toolkit. They exit at startup if no GPU is found. On hosts without a GPU, use the cpu tag and omit --gpus=all.

Choose a model

Browse the model registry or run:

self ls-remote
self pull kev:0.5b

The registry pins every download by size and SHA-256.