Self-hosted home AI inference platform

Multi-backend, multi-slot, extensible. Built for AMD Strix Halo (iGPU + NPU).

hal0.dev · Install · Docs

hal0 is a polished, reliable inference platform for running LLMs at home — a one-command install for any modern Linux box, with model slots, an OpenAI-compatible API, a built-in dashboard, and a prewired chat UI.

It targets AMD "Strix Halo" APUs (Ryzen AI Max, gfx1151) but runs anywhere llama.cpp/llama-server runs. On Strix Halo, hal0 takes full advantage of the iGPU (ROCm + Vulkan) and the XDNA NPU via Foundry Local Manager — model swap, mixed-backend slots, and unified memory addressing all the way up to 124 GiB.

• One-shot install — curl … | bash lands a working, idempotent stack. Re-running converges.
• Slot lifecycle as a state machine — every inference workload (chat, embed, rerank, STT, TTS, image) is a single-flight, systemd-managed unit with a known port and health probe.
• Capability-first UX — flat slots are real; the dashboard groups them into user-facing capabilities (Embed / Voice / Image / NPU rollup) so config stays legible.
• Provider plurality — llama-server, vLLM, Foundry Local Manager (XDNA NPU), moonshine (STT), kokoro/VibeVoice (TTS), Stable-Diffusion-WebUI all behind one OpenAI-compatible gateway.
• No vendor lock-in — Apache-2.0, open registry, OpenAI-compatible API surface. Bring your own models from HF.

v0.1.0-alpha — shipping. The cosign-keyless-OIDC release pipeline is wired end-to-end (signed tarball + Fulcio cert + manifest, self-verified before publish), and the one-line install actually installs. Expect rough edges: APIs may shift across 0.1.x alpha tags, no upgrade compatibility promised yet. See hal0/PLAN.md for the path to v1.0.

• Website — hal0.dev
• Email — [email protected]
• Issues — file in the relevant repo above