Replicate

Run open-source models through a hosted API with one call

AI Models & PlatformsFree planOverseasβ˜…β˜…β˜…β˜…β˜† 4.0

What is Replicate?

Replicate hosts thousands of open-source models and exposes each one through a simple API call, so you can run image, video, audio and language models without provisioning a GPU. You pick a model, send inputs, and either wait for the result or subscribe to a stream. The platform also lets you package your own model with Cog and deploy it the same way. The differentiator is breadth plus zero setup: it is the fastest path from reading about a model to getting real output, and per-second billing means idle capacity costs nothing.

Last updated: 2026-09-20. This site only provides an index; for exact features, pricing, and licensing, see the official website.

Key features

  • Hosted catalogue of image, video, audio and language models
  • REST API plus official Python, Node and HTTP clients
  • Cog packaging to publish your own model as an endpoint
  • Per-second GPU billing with automatic scale to zero
  • Fine-tuning support for several image and language models
  • Webhooks and prediction polling for long-running jobs

Pros & cons

Strengths

  • No GPU setup or driver work before your first prediction
  • Huge model catalogue makes experimentation cheap
  • Scale to zero means idle endpoints cost nothing

Watch out for

  • Cold starts add delay to infrequently used models
  • Heavy sustained workloads can cost more than owning GPUs
  • Popular models get crowded, so queue times vary

Best for & use cases

image and video generation, audio processing, rapid prototyping, custom model hosting

If you're comparing similar products, check the alternatives below, or browse all tools in the AI Models & Platforms category.

FAQ

How does Replicate pricing work?

You pay for the compute time a prediction uses, billed by the second, and idle models cost nothing because they scale to zero. Costs therefore depend on model size and how long each run takes.

Can I deploy my own model?

Yes. You package it with Cog, push it to a registry, and Replicate builds and serves it with the same API as the hosted catalogue. Hardware type and scaling options are set at deploy time.

Why is the first request sometimes slow?

Popular models stay warm, but less-used ones have to load onto a GPU first, which adds a cold-start delay. Keeping a model warm costs more, so the trade-off is latency against spend.