'hal0' local LLM server for Strix Halo mini-PCs enters public beta
hal0.dev is a local homelab inference server built specifically for hardware (mini-PCs with a Radeon integrated GPU, an XDNA NPU, and one pool), launching public beta this weekend. It exposes every AI modality through a single endpoint (:8080/v1): chat, embeddings, rerank, , , and . Under the hood it runs llama.cpp (with Vulkan, ROCm, FP4, and MTP support) alongside FastFlowLM for the XDNA NPU, running both simultaneously to squeeze out more performance, while is handled through ComfyUI.
A single install command is designed to wire up the entire stack automatically. The creator says the philosophy was deliberately narrow: bundle proven existing tools rather than reinvent them, with unified-memory usage and kanban/agent integration as the core focus.
Key points
- Local LLM server purpose-built for Strix Halo hardware (Radeon GPU + XDNA NPU + )
- Single covers chat, embeddings, rerank, STT, TTS, and
- Runs llama.cpp and FastFlowLM together to use both GPU and NPU
- handled via ComfyUI integration
- One-line install automates the whole stack; public beta opens this weekend