# Overview

> Develop, run notebooks, fine-tune and serve models on dedicated GPUs you control.

An instance is a container running on dedicated GPU or CPU compute. You connect to it through the browser terminal, SSH or public HTTPS ports.

Use an instance for work you want to interact with or keep up, such as:

- developing and debugging code on a GPU, from the browser terminal or your IDE over SSH;
- running notebooks in JupyterLab;
- fine-tuning a model while you watch and adjust it;
- serving a model behind an API with vLLM or SGLang;
- hosting a web app or demo, such as a Gradio app.

For work with a clear end that doesn't need you, such as a training run or a batch job, a [Run Job](/run-jobs) stops billing as soon as the command exits.

## Get started

<card-group>
<card icon="i-lucide-rocket" title="Deploy an instance" to="/instances/deploy">

Choose an image, compute and storage, and launch.

</card>

<card icon="i-lucide-plug" title="Connect to an instance" to="/instances/connect">

Open a terminal or connect with SSH.

</card>

<card icon="i-lucide-sliders-horizontal" title="Manage instances" to="/instances/manage">

Start, stop, restart, edit and delete.

</card>
</card-group>

## Concepts

### Storage

Each instance has a container disk for temporary files, which is erased when the instance stops or restarts. Mount [network volumes](/storage/network-volumes) for anything you need to keep. See [Storage](/storage).

### Connection

Every running instance has a browser terminal. Add an SSH key to connect with your own SSH client, and expose HTTP ports to reach web apps and APIs at public HTTPS addresses. See [Connect to an instance](/instances/connect).

## Deployment options

You can deploy an instance in several ways:

- **From an official template**: maintained environments such as PyTorch, vLLM and SGLang, ready to use.
- **From your own templates**: launch settings you or your organization saved, deployed the same way every time.
- **Custom images**: any Linux image from Docker Hub, GitHub Container Registry or another registry, public or private.

## How instances are billed

Instances bill per second for compute and container disk while the container runs. Stopped instances don't bill for compute. See [Billing](/account/billing).

## Limitations

- **HTTP only on public ports:** public ports carry HTTP, WebSocket and server-sent events. Raw TCP and UDP aren't supported.
- **Network volumes are set at deploy:** you can't add or remove a network volume on an existing instance.
- **Secrets are set at deploy:** you can change environment variables later, but not secret variables.
- **Linux images only:** images must support `linux/amd64`.

## Next steps

<card-group>
<card icon="i-lucide-rocket" title="Quickstart" to="/get-started/quickstart">

Deploy and connect to your first instance.

</card>

<card icon="i-lucide-notebook-pen" title="Develop in JupyterLab" to="/examples/jupyterlab">

A GPU notebook with your work on a network volume.

</card>
</card-group>
