documentation · versioned
everything, and where it lives
Reference for the version you actually run. Start at the top if the box is new; otherwise every section, and the whole knowledge base, is one click from here.
start here
bare box to first token
01Install hal0One line: fetch, verify, unpack, bring the API up.02Load your first modelPull from Hugging Face and assign it to a slot.03Send your first chatOpenWebUI, or curl the OpenAI-compatible endpoint.
docs
five sections, in reading order
Diátaxis, roughly: learn it, then apply it, then run it, then look it up.
knowledge base
things that stay true
Undated on purpose — every page carries a last-reviewed stamp instead, and gets revised rather than superseded.
getting started
Install, first slot, first request. The path from a bare box to a working /v1/ endpoint.
- Your first slot
- Talking to /v1/chat/completions · planned
- Picking a loadout · planned
hardware notes
What each machine can actually hold, and where the wall is.
- Unified memory is not VRAM
- Strix Halo: 128 GB, ~96 GB addressable · planned
- HX 370 on 32 GB · planned
runtime & backends
ROCm, Vulkan, the NPU lane, and how to tell which one you are on.
- ROCm vs vulkan_radv
- Building the rocmfp4 fork · planned
- XDNA and FastFlowLM today · planned
models & quants
Choosing a model that fits, and a quant that does not ruin it.
- Reading a quant name
- q4 vs q8 KV cache · planned
- MoE models on unified memory · planned
operating hal0
Running the box day to day: slots, eviction, logs, upgrades.
- Reading the memory map
- Co-resident slots without eviction · planned
- Upgrading without downtime · planned
tool reviews
Honest notes on the things people plug into hal0.
- OpenWebUI as the chat tab
- Continue.dev against a coder slot · planned
- MCP servers worth running · planned