# Dedicated endpoints

> Reserve GPUs for a single model with a private base URL.

A dedicated endpoint runs one model on GPUs reserved for you. It uses the same API as serverless inference, on a private
base URL.

## When to use it

- You need **consistent latency** at high, steady throughput.
- Your traffic exceeds serverless rate limits.
- You want a model we don't serve on serverless, or a specific precision.

## What you get

| | Serverless | Dedicated |
|---|---|---|
| Capacity | Shared, fair-use limits | Reserved GPUs, single tenant |
| Rate limits | Per-account defaults | Sized to your traffic |
| Billing | Per token | Per GPU-hour |
| API | `https://api.kurrens.ai/v1` | Private base URL, same API |
| Data retention | None | None |

## Getting started

Dedicated endpoints are provisioned with our team. [Tell us about your workload](/dedicated) — model,
expected tokens per day or peak requests per second, latency target, and region — Johor, Jakarta, or Dublin (see [Regions](/docs/getting-started/regions)) — and we'll size a
deployment. A dedicated endpoint stays in the region you choose.

Using it is the same as serverless — only the base URL changes:

```python
client = OpenAI(base_url="https://<your-endpoint>.kurrens.ai/v1", api_key="<KURRENS_API_KEY>")
```