Load your first model

Models in hal0 live in a local registry, the single source of truth for Hugging Face coordinates, SHA-256 digests, and curated filenames. You can populate it through the hal0 setup TUI or directly from the CLI.

Via hal0 setup

If you ran first-run setup and picked models for the Main and/or Agent slot, hal0 already seeded the registry and kicked off background pulls for each pick. Re-run hal0 setup any time to add more slots, models, or extensions; when a pull completes, its slot is ready to start. Nothing more is required.

Via the CLI

The Models view — registry, pull progress, and curated loadouts.

The hal0 model subcommand is a thin HTTP client to the API. The common flow is list → pull → assign.

1. List what’s available — registry rows plus models advertised by configured upstreams:

   hal0 model list

   Add --json to emit the raw /api/models payload for scripting.

2. Pull a model from Hugging Face. The argument is either a curated alias (e.g. qwen3-4b) or a registered model id:

   hal0 model pull qwen3-4b

   The pull runs as a background job on the daemon. The CLI polls /api/models/<id>/pull/status every 500 ms and prints a progress bar until the job reaches a terminal state, then reports the final path and SHA-256. If the curated pick is a vision model, its mmproj projector sidecar downloads automatically alongside the main GGUF — no extra step needed.

3. Assign the model to a slot’s default (this does not start the slot):

   hal0 model assign qwen3-4b --slot agent

   This writes model.default into the slot’s configuration. agent is the always-seeded default/anchor slot (see Slots); swap in utility or any other slot name you’ve created instead. Start the slot afterward (see Manage slots).

Other hal0 model commands: register (record an on-disk file in the registry), show (inspect metadata), and rm (remove a registry row).

Gated or large models

Pulls that need a Hugging Face token read HF_TOKEN from the API environment. The easiest path is the dashboard: Settings → Secrets → HuggingFace token applies live without a restart.

When a pull fails

A pull job ends in one of these states: queued, running, completed, failed, or cancelled. On failed, the CLI exits non-zero and prints the error message returned by the daemon; the registry row stays in place so you can retry.

# retry — an already queued/running pull for the same id is reused,
# not duplicated
hal0 model pull qwen3-4b

The dashboard’s Models view surfaces the same failure inline with a Retry button, and you can always pull later from there or via the CLI. Re-issuing a pull for a completed/failed/cancelled job replaces it; a pull already queued or running returns the existing job handle rather than starting a second download.

Next

Once a model is pulled and assigned, you’re ready to send your first chat.