Replicate

Replicate is a web and API platform for software engineers and teams that want to run, fine-tune, and deploy machine learning models in production without managing GPUs, servers, or ML infrastructure.

Key capabilities include:

• Run open-source models via API across images, video, audio, text, and chat, with a one-line call in Node, Python, or HTTP, plus a web Playground to compare and test models.
• Automatic scaling that scales up for traffic and scales down to zero when idle, so applications can handle demand without pre-provisioning.
• Fine-tuning and custom deployment; users can fine-tune supported models with their own data and deploy custom models packaged with Cog (open-source) to generate an API server.
• Operational visibility with logging, monitoring, metrics, and per-prediction logs for debugging.

Pricing: Pay-per-use, billed by the second for active compute (for example CPU $0.000100/sec, Nvidia T4 $0.000225/sec, Nvidia A100 80GB $0.001400/sec). Some models price per output (for example $0.06 per image for Flux 1.1 Pro Ultra).

Notable for combining a large catalog of working, production-ready model APIs with scale-to-zero economics, which makes it a practical alternative to self-hosting open-source models on custom GPU infrastructure.