Generative media on serverless GPUs
Run text-to-video models like Mochi behind an API without reserving a single GPU. Beam provisions H100s per job, saves the rendered video, and returns a hosted URL.
Run large video models on H100s per job
H100s per job
Video models need serious VRAM. Grab an H100 for the render and release it the moment the file is written.
Explore docsWeights staged on volumes
Multi-gigabyte pipelines load from a persistent volume instead of re-downloading on every boot. Storage is not billed yet. Planned: $0.0001 per GB-hour of data stored, with no free allowance.
Explore docsAsync by design
Renders take minutes, not milliseconds. Task queues accept the job instantly and return a task ID to track.
Explore docsWebhook when rendered
Attach a callback_url and your app gets a POST the moment the video is ready, so you don't have to poll.
Explore docsHosted video outputs
Save the MP4 with Output and hand back a public URL straight from the API.
Explore docsPer-second economics
An H100 for a two-minute render costs two minutes of H100.
Explore docsShip an AI video API
Upload your weights
Download the model to a persistent volume once, and every container mounts it instantly afterward.
Deploy the pipeline
Wrap the diffusers pipeline in an endpoint or task queue with an H100 attached.
Host as HTTPS endpoint
POST a prompt and get a hosted video URL back, synchronously for short clips or via webhook for long renders.
# 1. Stage the weights on a volume (once)
$ python upload.py # snapshot_download() into the volume
# 2. Deploy the generation API
$ beam deploy app.py:generate_video
=> Deployed 🎉
=> https://mochi-1-abc123.app.beam.cloud
# 3. Render over HTTP
$ curl -X POST https://mochi-1-abc123.app.beam.cloud \
-H 'Authorization: Bearer YOUR_TOKEN' \
-d '{"prompt": "a hot air balloon drifting over glaciers"}'
{"url": "https://app.beam.cloud/output/id/77c1..."}Frequently asked questions
Which video models can I run?
Beam has a step-by-step Mochi-1 guide, and any open video model that runs in PyTorch (via diffusers or custom code) deploys the same way.
How do you handle multi-minute render times?
Run the model behind a task queue: requests return a task ID instantly, timeouts are configurable (or disabled), and a callback URL notifies your app when the render lands.
How do I keep costs under control?
Billing is per-second and containers scale to zero between renders, so you pay for render time only.
Where do model weights live?
On a persistent volume. Upload them once with snapshot_download, and every container loads from disk instead of pulling tens of gigabytes per boot.
How are finished videos delivered?
Save them with Output and return hosted public URLs from your API, or push them to your own storage from inside the container.
Related use cases
Run Stable Diffusion and Flux as an API
Host SDXL, Flux, and custom checkpoints behind an autoscaling API.
InferenceTurn ComfyUI workflows into production APIs
Turn ComfyUI workflows into autoscaling API endpoints.
Task QueuesFan out batch jobs across thousands of containers
Fan out batch inference and ETL across thousands of containers.
Give your agents their own machines
Sign up, install the CLI, and launch a sandbox, VM, or GPU from your terminal or your coding agent.