Models on disk

What it shows

The Models card on the Dashboard lists every GGUF file currently on disk in your Blueprint models directory:

• Display name + quant: e.g. "Qwen2.5 7B Instruct · Q4_K_M"
• File name of the actual .gguf
• Size on disk in human-readable bytes

When you're viewing a remote host (top-right host selector), the list is the GGUFs on that host's disk, not yours. The Pull-to-host button (see below) downloads onto the remote, not locally.

Where the files live

• Local: ~/.blueprint/models/ (Linux/macOS), %USERPROFILE%\.blueprint\models\ (Windows).
• Remote: ~/.blueprint/models/ on the remote host's home directory.

You can verify with your file manager or ls ~/.blueprint/models/. Blueprint doesn't move files outside this directory; deleting one from the file manager removes it from the Models card on next refresh.

How to add models

Locally

Use the Plan wizard to pick a model + quant, then Deploy → Pull model. The kernel downloads the GGUF from HuggingFace and drops it into ~/.blueprint/models/.

To a remote host

1. Connect to the host via Hosts tab.
2. Switch the host selector (top-right) to that host.
3. Open Dashboard → Models. Click Pull to host.
4. Pick a model + quant from the dropdown and click Pull. The download runs on the remote — you can disconnect and reconnect; it keeps going.

The progress row at the top of the card shows bytes downloaded / total / bytes-per-second. Pulls that fail (network drop, disk full) show a red banner with the error.

Resuming an interrupted download

The kernel writes to a .part file during download and renames to .gguf only on success. If a download dies mid-way (network drop, process kill), the next pull of the same model + quant resumes from where it stopped — no need to start over.

If the partial file is corrupted or you want to force a fresh start, delete the .part file manually from the models directory before re-pulling.

Removing models

From the Maintain tab, the Delete model action removes the file from disk and updates the card. There's no undo — if you delete a 30GB model and change your mind, you'll need to re-pull.

Common pitfalls

• Disk full: GGUFs are large (Q4 of a 7B is ~4GB; Q4 of a 70B is ~40GB). The Models card sums total bytes — watch that against your free space, especially on remote hosts where you might not have a desktop file manager in front of you.
• Catalog ID vs file name mismatch: if you manually copied a GGUF into the models directory, the parser tries to guess the catalog ID from the filename. Names that don't match the pattern <model-id>-<QUANT>.gguf may show as the raw filename instead of the catalog display name. This is cosmetic — the model still loads.
• Pull to host on a host without enough disk: the pull starts but fails partway with a disk-full error. Check disk space on the remote before starting big pulls; the Overview tab shows free disk under the system snapshot.