Replicate is a cloud API for running machine learning models. Developers can call models from a community library with one line of code, fine-tune them on their own data, or deploy their own custom models.
The model library covers image generation, speech generation, music generation, image restoration, video generation, large language models and image captioning. Each model runs behind an API, so an application can send an input such as a prompt and receive the output without managing servers.
Teams that need more than an existing model can train a new one. The site describes fine-tuning image models such as SDXL on your own images so they produce a particular person, object or style. The resulting model is run the same way as any other model on the platform.
For custom code, Replicate uses Cog, an open-source tool for packaging machine learning models. Cog generates an API server and deploys it to a cloud cluster. The site says Replicate scales up to meet demand and scales down to zero when there is no traffic.
The product is aimed at businesses building AI features into their products. The site says a team can deploy an AI feature in a day and scale it to many users without being machine learning experts, and it offers enterprise plans.
Pricing is pay as you go. Public models are billed by the time they take to run, or by input and output for some models. Private models run on dedicated hardware and are billed for the time their instances are online, including setup and idle time. Hardware is billed per second, from a small CPU at $0.000025 per second to an Nvidia H100 GPU at $0.001525 per second.
