Modal

Modal runs your Python code in the cloud, on GPUs, without you building an image, writing a Dockerfile, or provisioning a cluster. You decorate a function, specify the hardware and dependencies in Python, and call it.

What makes it distinctive

• Sub-second cold starts on GPU containers, achieved through a custom container runtime. This is the technical achievement that makes serverless GPU inference viable rather than theoretical.
• Infrastructure as Python. Images, secrets, volumes, schedules, and hardware are declared in the same file as your code.
• Scale to zero and to thousands. You pay per second of compute actually used, which suits bursty inference and periodic batch work far better than a reserved instance.
• Web endpoints with a decorator, turning any function into an HTTP API.
• Scheduled jobs and queues for pipelines, plus sandboxes for running untrusted or agent-generated code.

Typical workloads

Model inference, fine-tuning runs, batch data processing, audio and video transcoding, and agent sandboxes — anything that needs a GPU occasionally and should not pay for one continuously.

Free monthly compute credits for getting started, then per-second usage-based pricing.