Skip to main content
The Cloud Inference page registers model ids that Belvedir executes on an endpoint of yours instead of its own account: a fine-tuned model you deployed, a self-hosted vLLM server, or a model on a provider account you hold. Calls to a registered id run on your endpoint with your key, metered for the usage graphs but never billed by Belvedir.

​
Register an endpoint

Enter the model id, the endpoint’s base URL, and its API key. The base URL must be https, publicly resolvable, and not a Belvedir host; a 3xx answer is not followed (502, gateway_redirect), so register the final host. If the endpoint knows the model under a different id (a Together deploy’s generated name, or a provider model you want to serve under a name of your own), fill in Model id it runs as: requests go out with that id, while routers, the Playground and the usage graphs keep the name you registered. Leave it blank when the ids match. Leave the base URL blank instead, and Model id it runs as names a hosted model Belvedir serves: your name then runs on Belvedir’s account as that model, billed at its rate like any hosted call, with no key of yours stored.

​
Where registered models appear

A registered id is accepted by the chat-completions endpoint, can be placed in any router tier or task-group assignment on the Routers page, and is listed in the Playground. Without a registration, local and fine-tuned model ids return 400 on the inference endpoints; the decision endpoint still routes to them for calls your own code runs. Models fine-tuned on Belvedir can be tested and deployed from the Models page before you register anything; see LoRA finetuning.