Skip to main content

User Models

User Models is the fourth tab of the LLM Models page: the models that already live on your own devices, registered with Cordatus so they can be launched without downloading them again — or transferred to another device that needs them.

Model paths on the device, scanning and importing, editing metadata, and the two launch paths

 

1. Configure the model paths on the device​

Cordatus only looks where you tell it to. Model paths are per device.

  1. Go to Devices and open the device.
  2. Open its Models tab.
  3. Under Storage Paths, add a path for each source you use:
    • HuggingFace cache path
    • Ollama models path
    • NVIDIA NIM cache path
    • Custom paths for anything else
  4. Downloaded Models then lists what was found under each path, with a Deploy button per model and a Refresh models action.

Removing a path only removes it from settings — your model files are not deleted.

note

Until a path is configured, that source reports "No HuggingFace path configured" (and the same for Ollama and NIM) rather than an empty model list, so you can tell "nothing there" from "nowhere to look".

caution

Model transfer needs a path on the destination device too. If it is missing you get "{type} model path is not configured on the destination device. Please configure it in Device Settings → Model Paths before transferring."

2. Import the models into your library​

Two ways, both from LLM Models → User Models.

Explore Models on Your Devices (scanning)​

  1. Press Explore Models on Your Devices.
  2. Pick a connected device — "Pick a connected device to discover the models it already stores."
  3. Start scanning. Cordatus asks the device for its HuggingFace, Ollama and NVIDIA NIM model directories, then reads each one.
  4. The result lists every model found, with its size and type. Models already in your library are marked Added; the rest are New. Folder sizes are calculated as it goes.
  5. Choose Target engines for the models you are importing.
  6. Add {n} model(s) imports them.

Rescan re-reads the device and cleans up anything it no longer stores.

note

GGUF weights run on llama.cpp only, which the dialog states while you pick target engines.

Add a model manually​

For a model in a path the scan does not cover:

  1. Add a model manually (or Add Custom Model).
  2. Select Device — a connected device.
  3. Model Path — browse to the directory.
  4. Model Name — taken from the path by default, editable.
  5. Select Engine — the inference engine(s) this model is for.
  6. For Ollama models Cordatus detects the tag and asks which to use ("Select model tag (e.g., 1b, 4b, latest)").
  7. Save Model.

3. The User Models list​

Each entry shows:

FieldMeaning
ModelIts name.
DeviceWhich device stores it.
PathWhere on that device.
SizeOn-disk size.
ParametersParameter count, when known.
QuantizationThe quantization, when known.
CreatedWhen it was registered.

Filter the list with Filter by engine, and select several rows for Delete Selected.

Editing a model​

Edit opens Update model metadata across all engines:

  • Model Name — fixed; the name cannot be changed.
  • Model Size, Parameters (format: 7B, 500M, 1.5B), Quantization (with its bit width shown).

The dialog states how many engine registrations the change will be applied to.

4. Launching a User Models entry​

Two buttons per row:

  • AI Launch — "Plan the engine, image, GPUs and flags for this model on this device." Because the device is already known, AI Launch plans for that device directly.
  • Manual → Run manually with… — opens the launch wizard with the model pre-selected.

Inside the wizard, User Models is also one of the three model sources in the Model pane, with a per-model status:

  • Complete – Ready to use
  • Incomplete – Can resume transfer
  • Transfer required

5. Model transfer between devices​

If you select a model that lives on a different device from the one you are deploying to, Cordatus offers to transfer it.

DialogWhen
Model Transfer RequiredNormal case. Shows the model size and the source path.
Resume Model TransferAn earlier attempt was incomplete. Shows how much is already there, and how many files are missing or only partially downloaded.
Corrupted Files DetectedChecksum verification failed; the affected files are re-downloaded.

During the transfer you see the overall percentage, bytes downloaded of total, files completed of total, and a per-file progress bar for whatever is in flight. Cancel Transfer stops it; the partial download can be resumed later.

Use API Only skips the transfer entirely when the model is reachable over an API instead of being copied.

When it finishes, the volume mapping for the model is configured automatically for the engine you are launching.

tip

Transfers run device-to-device over your own network, not through your browser. You can close the dialog; the launch waits for the transfer to complete.

Best practices​

  • Register the same three paths on every device you deploy to. It is what makes transfer, scanning and "already on device" detection work.
  • Keep parameter count and quantization filled in — the GPU fit advisor and the VRAM calculations use them.
  • Use Rescan after you delete models on the device, so the library does not keep entries for weights that are gone.
  • Prefer AI Launch from a User Models row: the model is local, so the plan is made from the actual files rather than from Hugging Face metadata.