Runtime Families

Runtime families describe the shape of work AI Cloud can launch and operate.

Overview

Runtime family	Best for	Typical connect surface
Compute	Raw GPU node or slice access	SSH, browser terminal, metrics
Scheduler app	Multi-node or batch jobs	Scheduler UI, terminal, job tooling
Compose app	Containerized services	App route, ports, logs, health
Notebook or IDE	Interactive development	Jupyter, VS Code, browser route
Inference endpoint	Model serving	API endpoint, model health, route

Compute

Use compute when you need direct access to capacity. Compute is the closest model to a raw GPU node or GPU slice.

Check before launch:

• SKU or slice profile;
• image;
• region;
• SSH or terminal access;
• hourly burn;
• release path.

For a step-by-step compute walkthrough, use Launch Compute.

Scheduler Apps

Use scheduler apps for jobs that need a scheduler, queue, or multi-step execution model. Examples include Slurm-like and Kubernetes-style workflows.

Confirm:

• scheduler UI or CLI access;
• job submission path;
• storage requirements;
• cleanup behavior;
• quota and team ownership.

Compose Apps

Use compose apps for container stacks with declared ports and health checks.

Confirm:

• exposed routes and ports;
• health checks;
• logs;
• storage references;
• release and cleanup behavior.

Notebook And IDE Apps

Use notebooks and IDEs for interactive development, experiments, data inspection, and debugging.

Confirm:

• route opens in browser;
• workspace persistence expectations;
• project storage;
• token and session behavior;
• stop/release behavior.

For app packaging and endpoint declaration details, use App Developer Guide.

Inference Endpoints

Use inference endpoints for model serving and OpenAI-compatible API patterns.

Confirm:

• model artifact or registry reference;
• endpoint route;
• authentication boundary;
• model health;
• scaling and quota expectations;
• release and cleanup behavior.

Inference endpoints should publish API-oriented routes and health checks. The app manifest should declare the endpoint as an API route rather than a browser workspace.