Sharing a model

A model whose weights live on one of your shared filesystems can be published for anyone on EcoLink to run. They launch their own instance against your weights; you keep paying for the storage and nothing more.

What it costs

Publishing changes your bill by nothing. Shared filesystem storage is charged by capacity, not by who reads it — the same amount whether nobody mounts your model or fifty accounts do.

Readers pay for the GPUs they run and nothing for the storage.

Publishing

You need the weights on a shared filesystem, in their own directory.

1. Put the weights in a directory of their own. Readers see that directory and nothing else on the volume. /mnt/shared/models/my-model works; /mnt/shared — the filesystem's own root — is refused when you register the model:

   model_path "/mnt/shared" is the filesystem's root. Put the weights in a directory of their own (/mnt/shared/my-model) — a model published from the root cannot be mounted by other accounts

   A root path works fine for you — the whole volume mounts and the weights are there — which is why it used to be allowed. It cannot work for anyone else: publishing grants access to the model's own directory, and a root path leaves nothing to narrow the mount down to.

   The same rule applies after registration. You cannot re-point a model at the root, and you cannot make one public while its path is already there. A model registered at the root before this check existed can still have its description edited — it just has to move into a subdirectory before it can be published.

2. Register the model (Registering a model), with Shared filesystem as the source and that directory as the path.

3. Set its visibility to public on the model's page in the Registry.

That is all. There is no separate step for the filesystem: publishing the model is what grants read access to its directory.

Give it a startup command

Set the model's default startup command before publishing. It is the only instruction a reader gets for how to run your model — without it they are guessing at arguments. It appears pre-filled in their launch form.

Running someone else's model

Registry → Models → Community lists every model other accounts have published, with its region. Each card offers two ways to run it:

• Launch as model instance: a managed endpoint serving the model (User Inference).
• Launch as GPU instance: a GPU machine with the weights mounted, where you have a shell, SSH or Jupyter and run the model however you like. Shown for models whose weights are on a shared filesystem.

You can also start from the launch forms:

• In Model Instances → Deploy, the model picker has three sources. Choose Community.
• In GPU Instances → Launch Instance, choose the model under Model weights.

Each row shows the model's name and its region. Pick one and the publisher's startup command fills in automatically, if the field is empty.

The region has to match

A model can only be run in the region its weights live in. A model published in mv cannot back an instance elsewhere — the weights are not there, and there is no copy.

The region is shown on every row so you can see this before you start rather than after.

You get read-only access

You can read the weights and run the model. You cannot change or delete anything, and you see only the model's own directory — not the rest of the publisher's filesystem.

Un-publishing

Set the model back to private. New launches stop immediately.

Instances already running keep their mount until they stop. This is deliberate: cutting a mount out from under a running job would destroy work that is halfway done, and the person running it has no chance to save anything first.

warning

Un-publishing does not undo publishing. It stops the next launch; it does not take back what has already been read. Anyone who ran your model may have copied the weights. Publish accordingly.

Deleting a model

A model cannot be deleted while something still points at it:

• Compute templates that reference it, whatever state they are in. A template exists to be launched later, so there is no state in which the reference is harmless.
• Instances that are still live — pending, running or degraded. A stopped instance is a record of something that already happened and does not block.

The error names the ids, so there is something to act on:

model "my-model" is still used by running instances 41, 42 and compute templates 7. Delete or re-point those first.

The message says "running instances" for all three live states, so an instance it names may be pending or degraded rather than actually running.

Deleting used to always succeed, and the references it left behind did not announce themselves: a template kept pointing at a model that no longer existed, and the launch form it opened simply had no model selected and said nothing about why.

Things worth knowing

A borrowed model runs on a model instance or a GPU instance, never on a cluster. Either way the reader names the model, and only its directory is mounted, read-only. Attaching another account's filesystem directly, by its id, is refused everywhere: without a model there would be nothing to narrow the mount down to a single directory, which is the whole point of sharing a model rather than a disk.

A GPU instance gives the reader a shell. They can read every file in your model's directory and copy it elsewhere, as the warning under Un-publishing says. Publish only weights you are happy to redistribute.

Your filesystem stays yours. Readers get the model's directory, read-only. They cannot browse the rest of the volume, and they cannot write anywhere on it.

If you suspend or delete the filesystem, the model stops working for everyone using it — including you. A suspended filesystem is unusable to its owner too.

Your account balance still matters. If it goes negative the filesystem is suspended, and every instance running your model loses its weights. Publishing a model is a small commitment to keeping it available.

Troubleshooting

"Model not found or not available to this account" — the model was un-published, or it was never public. Ask the publisher.

The instance will not start, and the region looks right — check the publisher's filesystem is not suspended. There is no way to see this from the outside; ask them.

A model has no region shown — it is HuggingFace-backed rather than filesystem-backed. Those are downloaded at launch and are not affected by any of the above.