Overview
Modal is a serverless compute platform aimed at engineers who want GPUs, distributed batch jobs, and web endpoints without managing a single VM. You write a Python function, decorate it with resource hints (GPU type, memory, concurrency), and Modal handles scheduling, autoscaling, and teardown - you pay only for the seconds your code actually runs. In our hands-on use, the standout is the programming model: a training job, an inference API, and a scheduled scraper can all live in one script and share the same deploy flow. Because it autoscales to zero, a rarely used endpoint costs nothing between calls, which is hard to match on a reserved-instance cloud. It shines for spiky ML workloads, inference serving, and data pipelines, and it exposes raw containers when you need custom images. The trade-off is lock-in: workloads are expressed in Modal’s decorator API rather than portable Kubernetes manifests, so moving off later means rewriting.
Key Features
- Per-second billing with autoscale-to-zero - no cost for idle resources
- Wide GPU menu from T4 to B200/H100, allocated by a single decorator argument
- Distributed map and queue primitives for fan-out batch and pipeline jobs
- Web endpoints and cron schedules deployed from the same Python script
- Custom container images and secrets for reproducible, production-grade runs
- Volumes and observability (metrics, logs, rollbacks) included per workspace
Pricing
| Plan | Price | For | Notes |
|---|---|---|---|
| Starter | $0 | Indie devs and small teams | $30/mo free credits, 3 seats, 100 containers, 10 GPU concurrency |
| Team | $250/mo + usage | Startups scaling up | $100/mo credits, 5000 containers, 50 GPU concurrency, custom domains |
| Enterprise | Custom | Security- and scale-critical orgs | Volume discounts, SSO, audit logs, HIPAA, private Slack support |
| On-demand compute | Pay-as-you-go | Any workload | H100 ~$0.0011/s, A100 80GB ~$0.00069/s, CPU $0.0000131/core/s |
Comparison
Against Together AI and Replicate, which are model-serving platforms, Modal is general compute: you bring any code or model rather than calling a hosted inference endpoint. Versus fal.ai, which optimizes media inference (image/video) behind fast APIs, Modal is broader but leaves more to you. Pick Modal when you need arbitrary Python jobs, training, or web services on GPUs; pick the others when you just want a specific model served with minimal code.