Container storage is ephemeral, anything written inside the container is gone when the service ends. Attach a persistent storage bucket if you want weights or files to survive a relaunch.

Only if you skip the storage bucket. Without one, the app re-downloads its full model set inside your paid session on every launch. With a bucket attached, it caches weights and reuses them.

No, endpoints are unauthenticated port forwards. Anyone with the URL can call your model, or for a notebook service, execute code in your container. Treat the URL as a secret.

No. Once you launch, your browser talks to the node operator directly over a peer-to-peer connection, and payment settles on-chain through escrow, there's no broker sitting between you and the hardware. (The Curated Models, Services, and Templates catalogues you browse beforehand are currently served from a single node as the network scales out, but each job you run still executes and settles directly between you and that node.)

The node operator hosting your container can see the traffic it handles and the data it processes, you're running on their hardware. For data that needs to stay private, Compute-to-Data is the better fit than Inference.

Yes, Edit lets you swap the model, engine, or configuration in place. The service ID, host port, endpoint URL, and remaining paid time all carry over. Changing resources or duration requires a fresh launch.