Built-in slots

hal0 always ships four slots out of the box. They live in BUILTIN_SLOTS (src/hal0/slots/manager.py) and cannot be deleted from the dashboard — you can swap their model, unload them, or leave them offline, but the slot itself is always present.

The four slots

Slot	What it serves	Default backend
primary	Chat and general LLM — /v1/chat/completions, /v1/completions	llama.cpp (Vulkan)
embed	Embeddings — /v1/embeddings — and rerank — /v1/rerankings	llama.cpp (Vulkan)
stt	Speech-to-text — /v1/audio/transcriptions	Moonshine
tts	Text-to-speech — /v1/audio/speech	Kokoro

Why four

These map directly to the modalities OpenAI exposes through /v1/*. Any client written against the OpenAI SDK can hit hal0 unmodified and reach chat, embeddings, transcription, and speech. Rerank piggybacks on the embed slot because it uses the same backend process.

Addressing

All four bind to 127.0.0.1 on a port in the slot range (8081–8099). Only the API (:8080) and OpenWebUI (:3001) bind public interfaces. Clients should always talk to the API, never to a slot directly — the API does authentication, single-flight, and structured-error wrapping.

You address a slot by its name in the OpenAI model field:

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "primary",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

The dispatcher resolves "primary" to whichever model is currently loaded in the primary slot. See Slot as model for the full convention.

Swapping the default model

Every slot has a default model picked at install time by the hardware probe. You can swap it at any time:

hal0 slot swap primary --model qwen3-30b-a3b-instruct-2507-q4_k_m

The slot transitions through unloading → starting → warming → ready without dropping the API socket — in-flight requests on other slots keep flowing.

Note

See Recommended loadouts for curated primary / embed / stt / tts picks at every memory tier.

User-defined slots

Beyond the four built-ins, you can add custom slots — e.g. a vision slot for a multimodal model, or an npu slot for the FLM provider on AMD XDNA hardware. See Custom slots.