Load your first model

A fresh hal0 install boots into the FirstRun wizard. Three steps, plus a “done” coda. Open the dashboard at http://localhost:8080 and the wizard takes over the screen until the primary slot has a model.

What the wizard does

The wizard is a guarded route at /firstrun. Until the API reports first_run: false, every other dashboard navigation redirects back to it — there’s no point operating an empty box.

It owns three jobs: pick a starting model from the curated list, confirm any license requirements, and stream a model pull straight into the primary slot. When the slot transitions to ready, the wizard moves to the done step and unlocks the rest of the dashboard.

The three steps

1. Pick model. A curated list of starting picks, sized to your detected hardware. The probe already wrote /etc/hal0/hardware.json during install, so the list shows fit warnings inline — “this model is larger than your detected GPU” appears next to anything that would offload heavily.

   The default highlight is Phi-3-mini-4k-instruct-q4 — a 2.4 GB Q4 GGUF that downloads in roughly 10 seconds on a modern connection and is small enough to fit anywhere hal0 runs. Strix Halo users should jump straight to a Q4 7B-class model or a Q4 MoE 30B — see recommended loadouts on the Strix Halo page.

2. License. Some weights have a click-through license (Llama, Gemma, others). When the picked model requires it, step 2 shows the license URL and asks for explicit confirmation. If the model is freely redistributable, this step is skipped.

3. Install. The dashboard kicks off a streaming pull through POST /api/models/{id}/pull. Progress streams live over SSE — bytes, percent, and the slot state transitioning through pulling → starting → warming → ready. The slot starts automatically when the download completes.

When primary reaches ready, the wizard moves to the done step. From there you can head straight to OpenWebUI for a first chat, or to the dashboard to load more models.

What if you want to skip the wizard?

The CLI gets you to the same place without a browser:

hal0 slot load primary --model phi-3-mini-4k-instruct-q4

The slot manager handles the pull, the systemd unit, and the health probe. hal0 status shows the same state machine the wizard streams.

When the pull fails

The wizard surfaces errors inline. Common ones:

• No disk space. /var/lib/hal0/models/ ran out mid-pull. Free up space and retry — partial downloads resume.
• Hugging Face rate-limit. Anonymous pulls hit a rate cap on popular weights. Export HF_TOKEN (or set it in /etc/hal0/api.env) and retry.
• License not accepted on Hugging Face. Some gated models require acceptance on the HF side before the API will serve the files. The error message links out to the model page.

The slot stays in the error state with details in /var/lib/hal0/slots/primary/state.json until you retry — nothing hidden.

Picking something other than the default

The wizard’s curated list is a starting point, not an exhaustive catalog. After it’s done you can:

• Add models from the Models page in the dashboard

• Assign them to slots from the Slots page

• Or do it from the CLI:

  hal0 model list
  hal0 slot swap primary --model qwen2.5-coder-7b-instruct-q4_k_m

See recommended loadouts for a hardware-by-hardware breakdown of what fits where.

Status — `model pull` (CLI) is staged

hal0 model pull returns NOT_IMPLEMENTED in the current build; the streaming pull goes through the API + dashboard for now. The wizard covers the same path end-to-end and is the supported flow.

Next steps

Send your first chat OpenWebUI is already pointed at your primary slot.

Slot lifecycle The state machine you just watched the wizard drive.

API reference Hit `/v1/chat/completions` directly from any OpenAI SDK.