Templates determine what software is installed on your GPU instance at launch. Instead of setting up CUDA, PyTorch, or Jupyter yourself, pick a template and start working immediately.

​
Browse templates

Navigate to Templates in the sidebar to see all available templates, organized by category. Each template card shows:
  • Name and description
  • Tags — Frameworks and tools included (e.g., PyTorch, Jupyter, CUDA, Docker)
  • Deploy button — Jump straight to the deploy page with this template pre-selected

​
Available categories

​
PyTorch & ML

Templates with popular ML frameworks pre-installed:
  • PyTorch + Jupyter — PyTorch with JupyterLab, ready for training and experimentation
  • TensorFlow + Jupyter — TensorFlow with JupyterLab
These include CUDA drivers, cuDNN, and the framework’s GPU-enabled build.

​
Development

Minimal environments for custom setups:
  • Bare CUDA — NVIDIA CUDA toolkit only, no frameworks. Use this if you install your own stack.

​
System

Base images for advanced users who want full control over the software stack.

​
Template availability

The template selection may vary depending on the offer you choose. Some offers show a searchable catalog of recommended templates, while others show templates grouped by category. NVIDIA CUDA is a safe default if you’re not sure which to pick — it gives you GPU drivers and the CUDA toolkit with no extra frameworks.

​
Using templates during deployment

Templates are selected in step 1 of the deploy flow. The template you choose determines:
  1. What software is available when you SSH in
  2. Recommended disk size — Some templates suggest a minimum disk size for their dependencies
  3. Available ports — Jupyter-enabled templates expose a web interface alongside SSH
You can also reach the deploy page directly from a template card by clicking Deploy, which pre-selects that template.

​
OS images

Some offers let you select an OS image in addition to the template. The platform shows CUDA-enabled images sorted by version (newest first), so you can pick the CUDA version that matches your workload.