Self-hosted models overview

Self-hosted models let your tenant deploy and operate a model behind an OpenAI-compatible endpoint — the four values below carry you through the workflow.

The mental model

Term	What it means	Where you use it
Self-hosted models	The Console product area for deploying and operating models	Navigation and documentation
Model	The resource you deploy: one model source bound to compute and serving configuration	Status, Edit, Stop, Restart, and Delete actions
Endpoint	The tenant-level callable address shown on model detail pages	The client’s base_url after appending /v1
Served Model Name	The exact name that model exposes to clients	The request body’s model field

An endpoint is an address, not the model itself. Multiple models in one tenant can show the same endpoint; the request’s model value routes to a Served Model Name. Seeing the tenant endpoint on a model that is stopped or still provisioning does not, by itself, tell you whether the Served Model Name is callable: another ready model using the same name may still serve it.

Platform models or self-hosted models?

	Platform models	Self-hosted models
Who operates the model	OneNexus	Your tenant through the Console
Workspace and GPU quota	Not required to make a call	Required to deploy and run a model
Recommended API key	Platform	Inference authorizing the selected model’s Served Model Name; a Platform key cannot call tenant models
Address	Shared catalog endpoint	Tenant endpoint shown on the model detail page
Request model	Exact model value from the expanded row’s cURL request	Served Model Name

Use the Platform quickstart when you only need a first hosted-model response. Continue here when you need to self-host a public Hugging Face model or a finalized model version from your registry.

Choose the narrowest credential

A Platform key calls every published catalog model, including models added later. It cannot call tenant self-hosted models and cannot be narrowed. For an Inference key, selecting a model authorizes that model’s Served Model Name, not only that one resource. A replacement in the same tenant that continues using the same Served Model Name shares access; other Served Model Names and the shared catalog are refused. Use an Inference key for any workload that calls your own models, and always keep each host together with its generated model value.

Prerequisites

Before you open Deploy model, confirm the following:

• You can sign in to the OneNexus Console and select your tenant.
• The header shows the region where you want to deploy. Workspace and GPU quota are region-scoped.
• The tenant has a workspace. If the Self-hosted models page shows Set up your workspace, create it before continuing.
• Quotas shows an Inference GPU allocation in the selected region. Without it, GPU-backed actions such as Deploy, Edit, Stop, and Restart are read-only; Delete remains available so existing capacity can be released.
• You have one supported source:
  • a public Hugging Face repository in owner/name form; or
  • a finalized, servable model version under Models.
• Your role can deploy models and create API keys.

The region in the header is what scopes workspace and quota. Check it before anything else — the same tenant in another region has different capacity.

If workspace or quota is missing, use the action shown in the Console and complete the onboarding hand-off before continuing. Do not borrow another tenant’s workspace or key.

The first-deployment path

Follow in order:

1. Deploy your first model — choose a source, Served Model Name, Flavor, and replica count, then wait for Running.
2. Call the endpoint — create an Inference key and verify a real response.
3. Operate the model — scale, edit, observe, stop, restart, and clean up safely.
4. Troubleshoot self-hosted models — diagnose provisioning and request failures without guessing.

The readiness gate is always the same: proceed to client integration only when the model is Running, the detail page shows an endpoint, and Copy as cURL contains a complete request.

Current boundaries

• Each self-hosted model serves the one model source selected at creation. Edit does not replace that binding; deploy another model to serve a different one.
• The Open model flow accepts a public Hugging Face repository; a private or gated repository is not part of this documented flow.
• The Console exposes explicit replica counts. Taking a model offline is an explicit Stop action; do not assume traffic-driven autoscaling or automatic scale-to-zero.
• Serving configuration and supported request options can differ by model recipe. Start with the generated cURL before adding optional request fields or runtime overrides.

Your data is not retained

Prompts and outputs are not retained, and are not used to train or fine-tune OneNexus models. See Pilot & security.