fal is a serverless Python runtime for running generative-media models and custom inference code in the cloud. It hosts a gallery of image, video, 3D, and audio models behind HTTP endpoints, and it lets you deploy your own Python apps that scale up under load and down to zero when idle, without managing infrastructure.
Two Python packages cover the two sides of the platform. The fal package is the SDK and CLI for building serverless apps: you subclass fal.App, decorate methods with @fal.endpoint, run the app on fal with a temporary URL for testing (fal run), and publish it to a persistent endpoint (fal deploy). The fal-client package calls hosted model APIs or your deployed endpoints from Python; falclient.subscribe(modelid, arguments={...}) submits a request such as an image prompt to a model like fal-ai/flux/schnell and returns the result, and the same call works against your own app by passing its endpoint ID.
Features
- Model APIs: hosted image, video, 3D, and audio models callable by model ID with a FAL_KEY
- Serverless apps: fal.App classes with @fal.endpoint routes deployed as HTTP services
- Test then deploy: fal run gives a temporary URL; fal deploy creates a persistent endpoint
- Scale to zero: deployed apps consume no resources when they receive no requests
- Python client: fal_client.subscribe and related inference methods for models and custom endpoints
- Pipelines: chain models and code inside one deployed app
- Dashboard and API reference: web console for managing deployments alongside published API docs
