How to beginner · 3 min read

Modal spot instances explained

Quick answer

Modal spot instances are ephemeral GPU compute resources offered at a lower cost by Modal, ideal for interruptible AI workloads. Use the gpu parameter in @app.function to request spot instances, which run on spare capacity and may be preempted, enabling cost savings for batch or fault-tolerant tasks.

PREREQUISITES

Python 3.8+
Modal account and CLI installed
Modal token configured (modal login)
Basic Python knowledge

Setup

Install the modal Python package and log in to your Modal account to access spot instances.

bash

pip install modal
modal login

output

Successfully logged in as user@example.com
You can now deploy and run Modal functions.

Step by step

Define a Modal app and use the gpu="A10G-spot" parameter in the @app.function decorator to request a spot instance GPU. Spot instances are cheaper but can be interrupted, so use them for fault-tolerant workloads.

python

import modal

app = modal.App("spot-instance-demo")

@app.function(gpu="A10G-spot")
def run_inference(prompt: str) -> str:
    # Simulate AI workload
    return f"Processed prompt on spot instance: {prompt}"

if __name__ == "__main__":
    with app.run():
        result = run_inference.call("Hello from Modal spot instance")
        print(result)

output

Processed prompt on spot instance: Hello from Modal spot instance

Common variations

Use different GPU types like gpu="A100-spot" depending on availability. Combine spot instances with local caching or checkpointing to handle preemptions gracefully. Use @app.function(gpu="A10G") for on-demand instances if uninterrupted runtime is required.

Troubleshooting

If your spot instance is preempted, catch exceptions and retry the task. Ensure your Modal CLI is up to date to access the latest spot instance features. Check your Modal dashboard for spot instance availability and quotas.

✅

Key Takeaways

Modal spot instances provide cost-effective GPU compute by using interruptible spare capacity.
Use the gpu parameter with a spot instance type like "A10G-spot" in @app.function to request spot GPUs.
Spot instances may be preempted, so design your workloads to be fault-tolerant or checkpoint progress.
On-demand GPU instances are available by omitting "-spot" but cost more.
Keep your Modal CLI updated and monitor your usage via the Modal dashboard.

Verified 2026-04

Verify ↗