Skip to main content
When you create a Compute instance, you choose how it runs: as a **container **or as a virtual machine (VM). Both options can run real workloads. The difference is how much control you get over the operating system and how you manage software. If you’re unsure, start with a container. It’s usually the quickest path to getting work done, and you can switch by creating a new instance later if you run into OS limits.

​
Quick rule of thumb

Choose a container if you want a fast, managed runtime for scripts, apps, and models. Choose a virtual machine if you need full OS control, sudo access, kernel-level settings, or a specific Linux distribution.

​
Containers in Compute

Containers are designed for most common workloads. They’re a good fit when you want to run code and ship a working environment without managing the operating system. A container-based instance is usually the right choice when you’re:
  • Running scripts, apps, or APIs
  • Doing standard ML workflows with typical dependencies
  • Serving models and inference workloads with a supported template
  • Prioritizing a simple setup and quick startup
A few things to know up front:
  • Containers don’t support root access. Commands that require sudo won’t work.
  • System-level changes are handled through templates rather than modifying the OS directly.
Note: If your workflow depends on installing OS packages manually, running system services, or changing low-level settings, you’ll likely want a VM instead.

​
Virtual machines in Compute

Virtual machines are for workloads that need a traditional server experience. You get full control of the operating system, including the ability to install and configure software at the OS level. A VM is usually the right choice when you’re:
  • Installing system packages with your OS package manager and managing services
  • Running software that expects full OS behavior
  • Working with tools that require deeper system access
  • Needing a specific Linux distribution or custom OS image
  • Depending on kernel-level configuration or system drivers
With a VM, you can expect to work more like you would on a typical Linux server. That flexibility can be worth it, but it also means you’re taking on more responsibility for OS setup and maintenance.

​
Common scenarios

If you’re running a Python script, a small service, or a straightforward app, start with a container. If you’re serving an LLM using a standard inference setup, a container is often enough. Choose a VM if your setup requires OS-level customization, system services, or a specific distro. If you need to install system packages, configure services, or rely on sudo, choose a VM. If you’re not sure and you just want the simplest path to “it works,” choose a container.

​
What changes between containers and VMs

Here’s what users notice most in practice:
  • OS control: containers have limited OS control, VMs give full OS access
  • Installing software: containers rely on templates and app-level dependencies, VMs use standard OS package tools and configuration
  • Setup speed: containers tend to be faster to get running, VMs are more flexible but require more setup
  • Compatibility: containers cover most workflows, VMs are better when your workload depends on OS-specific behavior or a specific distribution
Tip: If you hit a wall with OS-level requirements, that’s a strong signal you should move the workload to a VM.

​
Switching if you picked the wrong one

You can switch approaches by creating a new instance of the other type and redeploying your workload. Before you switch, it helps to write down:
  • The template or base image you used
  • Your application dependencies and install steps
  • Any environment variables or secrets your workload needs
  • The commands you use to run your workload
If you’re moving from a container to a VM, plan for OS-level setup that you may not have needed before (packages, services, users, and startup scripts). [VERIFY]: If the product supports converting an existing instance from container to VM (or the reverse), document that here. If it doesn’t, keep the guidance as “create a new instance.”

​
Next steps