fal is a platform for generative media. It gives developers access to more than 1,500 image, video, audio and 3D models through a single API, and it lets them deploy their own models on serverless GPUs or rent GPU clusters for training.
The model library covers image generation and editing, text-to-video and image-to-video, text-to-speech and sound effects, image-to-3D, and world models. Each model is called through one unified API and SDK, so a team can add a new model without setting up a separate provider account.
For custom work, fal runs your own models or fine-tuned LoRAs on serverless infrastructure. The page says deployments autoscale across thousands of GPUs, with built-in observability and performance controls. Private model endpoints and bring-your-own-weights are also offered.
For training and heavy compute, fal offers GPU clusters with H100, H200, B200 and B300 hardware. Teams can fine-tune models on multi-node clusters and scale jobs across hundreds or thousands of GPUs.
The platform is aimed at developers and at enterprises that need SOC 2 compliance, single sign-on, private deployments, usage analytics and 24/7 priority support. Enterprise customers can also work with forward-deployed machine learning engineers.
Pricing is usage-based. Serverless models are billed per output, such as per second of video, per image, per 1,000 characters of speech, or per megapixel. GPU compute is billed per hour, with the RTX PRO 6000 listed from $1.99 per hour and the H100 from $2.49 per hour at the lowest tier. Reserved capacity is available through sales.
