Replicate is a cloud service that runs machine learning models through an API, an interface that another program calls over the internet. Its catalog holds thousands of open-source models contributed by the community, plus hosted proprietary models, for images, video, speech, music and text.
A program sends the model name and the inputs, such as a text prompt, in one call from the Python or Node.js client or over HTTP, and receives the output. The model runs on Replicate's GPUs. Models can also be fine-tuned, which means trained further on new data: a training job takes a zip file of images and produces a new model that can generate a particular person, object or style. Teams can deploy their own models with Cog, Replicate's open-source tool for packaging machine learning models; Cog generates an API server, and Replicate scales the model up and down with demand.
Replicate is a pay-per-use service. Most models are billed by the seconds they run, at a price that depends on the GPU type; some are billed per output, such as per image or per token.