Replicate

Replicate is the AI equivalent of an app store and serverless runtime in one. Any model published on Replicate (and many of the most-used image, video, and audio models are first-published here — FLUX, Stable Diffusion variants, Bria, Whisper, MusicGen, Stable Video Diffusion, Llama, plus thousands of community fine-tunes) can be called as a versioned HTTP endpoint without provisioning GPUs. Replicate handles autoscaling, queuing, cold-start optimization, and webhook delivery for long-running predictions. Developers can also push their own models using Cog, Replicate's open-source containerization tool, which turns a Python file + cog.yaml into a portable, GPU-ready prediction service that runs the same locally and in Replicate's cloud. Pricing is per-second of GPU time on the underlying hardware (with cheap, fast endpoints for popular models priced per-output — e.g., per image — to abstract the GPU detail away). For teams that need stability, Replicate offers Deployments (private, autoscaling endpoints with your own minimum-warm pool and scale rules) and Enterprise (dedicated capacity, SLAs, security review).