# fal A generative media platform for developers to call hosted models, deploy custom models, and access GPU compute. - Findings24: - Site: - Preços: Paid - Categorias: Plataformas de inferência de IA, Implantação de modelos de IA, Nuvem de GPU ## O que é fal? fal is a developer platform for building and operating generative media applications. It provides a unified API for hosted image, video, audio, music, speech, 3D, and multimodal models, alongside infrastructure for deploying custom models and running GPU workloads. Teams can start with a hosted model request, then use fal Serverless for custom endpoints or dedicated Compute instances for training and sustained workloads. The platform brings these workflows together so applications can use hosted and custom models through fal’s inference infrastructure. ## O que fal pode fazer? ### Unified model API Call more than 1,000 model endpoints for image, video, audio, and other media tasks through one API. Requests support synchronous and asynchronous queue calls; many models also offer streaming or real-time WebSocket connections. ### Custom serverless deployments Package a model as a Python app, test it on cloud GPU runners, and deploy an authenticated endpoint with autoscaling, retries, revision history, and rollback support. ### Dedicated GPU compute Provision GPU instances with SSH access for training, fine-tuning, batch jobs, and persistent workloads. Compute is billed hourly, while serverless inference is billed by execution time. ### Request monitoring and scaling controls The dashboard provides logs, request-level analytics, and error tracking. Deployment settings let teams configure minimum and maximum concurrency and pre-warming behavior. ### Model discovery and comparison Browse the model marketplace, review model-specific pricing, and compare models side by side in the Sandbox before selecting an endpoint. ### Usage-based pricing Hosted model APIs use per-output or other model-specific billing units, while serverless and dedicated GPU compute use GPU-based rates. Cost can vary with model settings such as resolution, duration, and quality. ## Casos de uso ### Add media generation to an application Use hosted endpoints to add image, video, audio, speech, or 3D generation without provisioning model infrastructure for each API request. ### Serve a proprietary or fine-tuned model Build a Python app around custom weights or a proprietary pipeline, test it on a cloud GPU runner, and deploy it as an authenticated endpoint that scales with demand. ### Train or fine-tune on dedicated GPUs Run training and fine-tuning jobs on dedicated instances with SSH access and fixed hourly billing, including multi-GPU workloads. ### Evaluate model options before integration Compare available models in the Sandbox and inspect their billing units and settings-dependent costs before choosing an endpoint. ## Perguntas frequentes ### How do I make a first model request? Choose a model from the marketplace, create an API key, and call its endpoint using the fal client libraries or the REST API. The documentation provides Python, JavaScript, and cURL examples. ### Can I deploy my own model? Yes. Define a Python fal.App, test it on a cloud GPU runner, and deploy it as a persistent endpoint. Deployments include authentication, autoscaling, retries, and revision-based rollbacks. ### When should I use Compute instead of Serverless? Compute is intended for training, fine-tuning, batch jobs, and persistent workloads requiring dedicated GPU access and SSH. Serverless is designed for API endpoints and on-demand inference with managed runners. ### How is usage billed? Model APIs use model-specific units such as outputs, seconds, characters, or megapixels. Serverless and Compute use GPU-based pricing; Compute is billed hourly and serverless execution by time used. Model settings can affect the final cost. ## Informações rápidas - **Product category**: Generative media inference and GPU compute platform - **Primary users**: Developers and teams building AI applications - **Model coverage**: Image, video, audio, music, speech, 3D, and multimodal models - **Custom deployment**: Python apps on managed serverless GPU runners - **Dedicated compute**: GPU instances with SSH access for training and persistent workloads - **Pricing model**: Model-specific API usage; serverless execution and dedicated GPUs billed by usage or hour