Point us at any OpenAI compatible Hugging Face model, open an app and set it up yourself, or start from a template that is already configured. The GPU is yours alone while it runs.
Curated packages pair a model with an engine preset already checked against the card it asks for, so nothing dies at startup. Anything else on the Hub goes through the custom flow, and both routes answer on the OpenAI API your code already calls.
Services are the bare apps. You get an empty ComfyUI, OpenWebUI, OpenCode or Hermes with nothing configured in it, then you go to Models, bring up the one you actually want, and set it up inside the app yourself.
Templates are those same apps with the setup already finished. We have chosen the models and built the workflow, so it runs the moment it starts.
The questions that come up every time somebody weighs us against what they use today.
| What you get | Ocean Network | Serverless APIs | Managed endpoints | Raw GPU rental |
|---|---|---|---|---|
| The model you name is the model that loads | Yes | Pooled, provider not pinned | Yes | Yes, if you build it |
| A card nobody else is sharing | Yes | No | Yes | Yes |
| Price agreed before anything starts | Hourly, shown upfront | Per token, after the fact | Hourly | Hourly |
| Weights kept between runs | Optional bucket | Not applicable | Re-pulled on cold start | If you wire it up |
| Running without configuring anything | Templates, services and curated models, one click | Yes | Some setup | You build the image |
You pay a node operator directly for the hardware and the minutes you use, and the window you book is a hard stop.
“If I cannot pin the model provider, I do not want it.”
Pick a package, an app, or any repository on the Hub, and pay for the window you booked. The docs walk through all three routes if you would rather read first.