Skip to content
HUIAHUB

Models & inference

Replicate

A cloud API for running, fine-tuning and deploying open-source and proprietary AI models, billed by usage.

Verified Paid
Visit official site (opens in a new tab)

What Replicate does

Replicate hosts machine learning models behind an HTTP API. You pick a model from its catalog, send inputs from Python, Node.js or plain HTTP, and get the output back. The catalog has thousands of community-contributed open-source models alongside proprietary ones from companies such as Google, OpenAI, Black Forest Labs and Anthropic, covering image, video, speech and music generation, image restoration and captioning, and large language models.

Beyond running public models, you can fine-tune some image models with your own data, and package your own model with Cog, Replicate's open-source tool for building production-ready containers. Private models and deployments run on dedicated hardware that you choose. Replicate scales them up and down with traffic. The docs also cover webhooks for finished predictions, streaming output and an MCP server.

Pricing is pay-as-you-go. Most public models bill by the second for the hardware they run on, from CPUs to Nvidia T4, L40S, A100, H100 and H200 GPUs, while others bill per output image, per second of video or per token. Each model's page shows a cost estimate. Private models bill for all the time an instance is online, idle time included, except fast-booting fine-tunes. Failed runs are not charged.

Replicate suits developers who want to add image, video, audio or language models to an app without running GPU servers themselves, and teams that want to host a custom model behind an API.

  • Generate images, video, speech or music from an app through one API.
  • Call hosted large language models without managing GPUs.
  • Fine-tune an image model on your own data.
  • Package a custom model with Cog and serve it on dedicated hardware.

Who should use it

  • Developers who want hosted AI models behind an API without running GPU infrastructure.
  • Teams deploying their own models who want autoscaling hardware.

Who should look elsewhere

  • Projects that must keep inference on their own hardware.
  • Teams that need a fixed monthly price, since billing depends on usage.

Strengths

  • Large catalog of open-source and proprietary models behind one API.
  • Cog, the packaging tool for custom models, is open source.
  • Per-second hardware billing for public models, with no charge for failed runs.
  • API-created prediction data is removed after an hour by default.

Limitations

  • Costs vary by model and hardware, so you need to check each model's page for estimates.
  • Private models bill for setup and idle time as well as active processing.
  • Free use is limited to select models for a short period before billing must be set up.
  • Multi-GPU and H200 capacity require committed spend contracts.

Privacy notes

Replicate's docs say inputs, outputs, files and logs from predictions created through the API are removed automatically after an hour by default, while predictions created on the website are kept indefinitely.

Sources