> See the full index of documentation pages at https://docs.zettabyte.com/llms.txt.

# Deploy a model

> Deploy a model as a dedicated inference endpoint.

Source: https://docs.zettabyte.com/docs/deploy-and-serve/deploy-a-model

Last updated: 2026-10-17

A deployed model runs as a dedicated inference endpoint. Endpoints appear under **Inference** in the zSTUDIO workspace, and also under **Tasks > Inference** in zCLOUD.

## Before you begin

- A model in **Library > Models**, either a preset from the catalog or one you brought yourself.
- GPU capacity for the endpoint's replicas. If capacity is short, you may need to **Join The Queue**.
- To check that a model fits in the GPUs you will request, see [GPU memory calculator for LLMs](https://docs.zettabyte.com/docs/deploy-and-serve/gpu-memory-calculator).
- An API key with the **Inference Service** permission, which you will need to call the endpoint afterward. See API keys.

## Deploy a model

From **Library > Models**, select **Deploy** on a model to open a five-step wizard with the model preselected. You can also start one from **Create Inference Task** on the Inference list, or use the simpler form under **Training and tasks**.

1. **Basic Information**: an **Endpoint Name** and a **Description**.
2. **Environment Configuration**: a **Distributed Inference** toggle, an **Inference Engine** (a preset such as `vLLM`, or a custom engine), the **Model** and version, and **Environment Variables**.
3. **Resource**: a **Computing Config** (**Resource Name**, GPU/CPU/Memory) with a **Complete Machine Scheduling** toggle, a **Number of Copies** (replicas), and **Storage**. Storage accepts a **Storage Volume** (PVC), an **Ephemeral Volume** (memory-backed, sized in GiB), and **S3** (access key, secret key, mount directory, region, and a path; an external AWS S3 URL is also accepted). Mount directories must be unique.
4. **Advanced**: optional **Scaling**. With scaling on, choose **Timed Scaling** (target replicas on a daily schedule, up to 10 rules) or **Automatic expansion and contraction** (target replicas plus a **Scalability Index** of **GPU Utilization Rate** with a target %).
5. **Overview**: review and create.

## Endpoint status

In the Inference list each endpoint shows Name/ID, Model Name, Inference Engine, Status, Scaling (ready/target replicas, for example `0/1` while starting), Creation Time, GPU Count, and Resource.

A new endpoint starts as Pending. As its replicas schedule and start, Scaling moves from `0/1` toward your target Number of Copies. The endpoint is ready to serve once Scaling shows every replica running, for example `1/1`.

You then call the endpoint over the inference API. The `Service-Name` header binds each request to your deployment and is built as `TENANT_ID.JOB_ALIAS_NAME`. It is a zCLOUD-specific (non-standard) header, so add it explicitly in every client. See API keys and Call your endpoint.

> [!NOTE]
> **Next step**
> A running endpoint isn't useful until you can call it. See Call your endpoint for paste-ready cURL and Python examples.

## Model Market

The Model Market is a display and consumption surface. Opening a listed model shows an **API and endpoint** tab and a **specs** tab, with a ready-to-copy cURL and Python snippet. There is no publish or create action on the Model Market; the only lifecycle control on a listing card is **Delist**. A model reaches the market from the deployment pipeline upstream, not from a create button on this page.