The inference platform for generative media and world models.

An inference platform runs trained AI models in production and serves their outputs over an API. It manages the GPUs, autoscaling, request queues and monitoring, so the team building the product does not have to.

AI inference is running a trained model on new inputs to produce outputs, such as generating an image from a prompt. Training builds the model once; inference runs every time a user makes a request, so its speed and cost decide the product's experience and margins.

fal runs world models through fal WMA, the World Model Accelerator. It streams causal and bidirectional diffusion models over WebRTC, scales from 1 to 1,000 GPUs across clouds and regions, and handles signaling and TURN relay. H3 Max Director and Abot World run on it today, and fal Worlds lets you browse them and start a live session.

fal is built specifically for generative media inference: an inference API for 1,400+ image, video, audio and 3D model endpoints, serverless inference for your own models with reserved capacity and burst, real-time WebSocket and WebRTC transport, and a marketplace to distribute the models you serve.

Serverless inference runs a model on GPUs that start when requests arrive and stop when they do not. On fal Serverless you set floors, caps and warm buffers per app, and you are billed per second while a runner is alive.

The inference API calls models fal already hosts, priced per output or per GPU second. Serverless inference runs your own model, pipeline or container on fal's GPUs, billed per second of runner time. Both use the same client, queue and calling patterns.

Yes. New models from leading labs go live on fal as they are released, with an API endpoint and a playground on day one.

Every account has a concurrency limit on how many requests run at once. New accounts start at 2 and the limit rises automatically with purchased credits, up to 40. Requests above the limit wait in the queue. Contact sales for higher limits.

Generated files are stored on the fal CDN for at least 7 days by default. You can set the retention period per request with the X-Fal-Object-Lifecycle-Preference header, and delete a request's stored payloads and output files through the Platform API.

Yes. A reservation guarantees a baseline of GPUs for your Serverless endpoints, and traffic above it bursts onto fal's shared fleet automatically. Contact sales to size one.

Write a fal.App class in Python with @fal.endpoint routes, bring a container image, or expose an existing HTTP server with exposed_port. Test with fal run and deploy with fal deploy.

Weights on /data are cached between startups automatically. FlashPack, fal's open-source loader, streams tensors from disk to GPU, and compiled kernels can be shared across runners. For latency-critical apps, min_concurrency and concurrency_buffer keep runners warm.

Inference API models are priced per output unit or, for some models, per GPU second, as listed on each model page, and time waiting in the queue is not billed. Serverless inference is billed per second while a runner is alive, by GPU type; time spent pending or pulling the image is not billed. Serverless GPU list prices per hour: B300 $12.99, GB200 $9.99, B200 $7.99, H200 $6.00, H100 $4.50, RTX PRO 6000 $4.00, with lower rates for volume and commitments. Reserved capacity is priced through sales. See all pricing

Yes, for supported models. Real-time endpoints keep a persistent WebSocket open to a runner, models with a streaming endpoint send outputs over server-sent events, and fal WMA streams over WebRTC for interactive applications.

Yes. Publish a Serverless app to the fal marketplace. Callers pay with their own fal key for their own usage, and the fal team configures unit pricing and the model card with you.

For large training runs, fal Compute provides dedicated GPU clusters on reserved terms. Contact sales to size one. You can deploy the trained weights to fal Serverless for inference.