LLM Models
LLM Launcher → LLM Models is the model catalogue. It brings four sources together — Hugging Face, Ollama, NVIDIA NIM and the models already stored on your own devices — and lets you filter them by the hardware you actually own, download them onto a device, and launch them.
The four tabs
Use the ‹ › navigator in the subheader to move between them.
| Tab | What it lists |
|---|---|
| Hugging Face | The Hugging Face hub, filtered and sorted for inference. |
| Ollama | The Ollama library, with the pull command for each tag. |
| NVIDIA NIM | NVIDIA NIM microservices. |
| User Models | Models already on your devices, registered with Cordatus. See User Models. |
Search
One field, which accepts either:
- free text — "Search models or paste URL…", or
- a full model URL, in which case a badge appears naming the source it was recognised as.
On the Hugging Face tab the field stays disabled until a search type is chosen.
Filters
Filters opens the filter panel; the button carries a count of how many are active.
Hardware filters
These two are mutually exclusive — Cordatus disables one while the other is set, and says why.
- Jetson Device — pick a Jetson module and the list narrows to what it can run. The parameter filter is then driven automatically by that device.
- GPU Device — pick a GPU model from the list, with a + / − stepper for how many of them you have. Each entry shows its VRAM.
- From Device — instead of naming a GPU, read the real GPUs off one of your connected devices. The dialog walks you through Choose a connected device → Select GPUs, showing total, used and free VRAM per GPU, and can be limited to Only devices with a GPU. Once applied, a chip shows the device and its total VRAM, with an × to clear it.
Catalogue filters (Hugging Face)
- Sort — by trending, downloads, likes, recency and so on.
- Parameter Range — a two-handle slider over the usual parameter sizes. Disabled while a Jetson device or a GPU is selected, because those drive it automatically.
- Advanced Filters — Engines, Publishers and Models, each dependent on the one before it (Select Engine First, Select Publisher First).
User Models filter
- Filter by engine — narrow the list to models registered for one inference engine.
Active filters appear as removable chips under the panel, with CLEAR ALL.
A model row
Each row shows the model, its publisher, its size and parameter count, and a lock indicator:
- "This model is open and does not require an access token."
- "This model requires an access token to download. You need to provide a valid Hugging Face token."
For Jetson, a Compatible Jetson Devices row lists which of your modules can run it, or says "No compatible Jetson devices found for this model." If the catalogue is missing the metadata needed to decide, it names what is missing rather than guessing.
Actions on a row:
| Action | Result |
|---|---|
| AI Launch | Opens AI Launch with this model. Find a device for this model and plan the engine, image, GPUs and flags. |
| Manual → Run manually with… | Opens the launch wizard with the engine you choose. |
| Deploy | Sends the model into the wizard for the engine whose catalogue you are browsing. |
| Download | Fetches the weights onto one of your devices. |
| Get Info | Fetches the model's details from Hugging Face when they are not cached yet. |
Load More LLM Models pages through the list.
Downloading a model onto a device
Download opens a dialog that does the whole job on the device, not through your browser.
- Device — every registered device, with an online/offline marker. If you have none: "No devices are registered to your account yet. Add one from the Devices page."
- Directory — the model directories registered on that device. If none is registered yet the dialog says so, and Add a directory registers one. When the directory needs to be created with elevated rights, Cordatus explains exactly what will be done: "Only the directory is created as root; the download itself is not."
- Quantization / Profile — when a repository ships the same model exported several ways, only
the one you pick is downloaded, plus the shared files. Every quantization downloads them all.
Loose
.gguffiles that are not a quantization of their own are excluded unless you choose Every quantization. - Token — choose a saved token or paste one. Optional for open models; optional, but faster where rate limits apply.
- sudo password, when the directory needs it. If a password is already stored on the device, Cordatus uses it and tells you: "It is never sent to your browser."
Besides the registered directories, Another directory… opens the device's own file browser, so a path that was never registered can still be used.
Before you start, the dialog states the cost — "About {size} per device will be fetched from the hub onto the selected device. Each device keeps going if you close this page." — and, when the weights turn out to be there already, it says so and changes the button to Verify on device: "Already here: all {n} file(s) are present. Starting anyway verifies them and downloads nothing."
While it prepares, the dialog narrates what it is doing — reading the device's directories, reading the repository's file list, checking whether the model is already there.
Problems are reported precisely rather than as a generic failure:
- "This model is already downloading on {device}. Starting again would write over the same files."
- "{path} is not writable by {user} — it belongs to another user."
- "The file list could not be read ({error}), so the whole repository will be downloaded."
A download in progress can be cancelled from the same button. Its percentage shows on the model's row, per device, and a counter appears in the header: Model downloads lists every job with its device, its state and the path it wrote to, and Clear finished empties the list.
Explore Models on Your Devices
On the User Models tab, Explore Models on Your Devices scans a connected device for models it already stores.
- Pick a device, then Start scanning. Cordatus asks the device for its HuggingFace, Ollama and NVIDIA NIM model directories and reads each one.
- Found models are listed with their size and type; ones already in your library are marked Added, the rest New.
- Choose Target engines for the models you import, then Add {n} model(s).
- Add a model manually covers a model in a custom path: name, path, type, tag and engines.
- Rescan re-reads the device; anything it no longer stores is cleaned up.
GGUF weights run on llama.cpp only, and the dialog says so while you pick target engines.
Full details: User Models.
Permissions
Without the model-management permission the launch and import actions are visible but disabled, with "You don't have permission to add models" on hover.