Modal (US) is a serverless compute platform aimed at AI workloads. Developers define Python functions, attach CPU, memory and GPU, and the platform runs the code on demand, exposing HTTP endpoints to applications. The focus is serving models and inference tasks without keeping servers running.
What it does
- Runs Python functions on demand with GPU
- Serves AI models through HTTP endpoints
- Runs batch processing jobs and scheduled tasks
- Scales inference with request volume
How it works
- Code is packaged with its declared dependencies
- Each function gets CPU, memory and GPU defined in the code
- Shared endpoints bill per token; dedicated endpoints bill by compute
- The platform starts and stops containers based on demand
- Endpoints, queues and persistent volumes are part of the runtime
Models and infrastructure
- Runs open models and user-provided custom models
- NVIDIA B300, B200, H200, H100, A100, RTX PRO 6000 and L40S GPUs
- Billed per second of CPU, memory and GPU used
- Monthly included compute on the Starter and Team plans
Public pricing
- Starter plan with no monthly fee and US$ 30/month of included compute
- Team plan from US$ 250/month, with US$ 100/month of included compute
- Per-second billing; H100 SXM5 around US$ 0.001097 per second
- Enterprise with volume discounts, SSO, audit logs and HIPAA
Strengths
- No charge for idle resources: you pay for execution time
- Python workflow without managing servers or Kubernetes
- Automatic scaling for traffic spikes
Points of attention
- Requires writing code on the platform or adapting the application
- Prices vary by region and by non-preemptible execution
- Corporate security features are on the Enterprise plan