Skip to content

Replicate

Run open-source models behind an API without managing GPUs

Pay-per-useModel APIs
50

Overview

Replicate is a cloud API platform for running and fine-tuning open-source machine learning models, built for developers who want to add AI features without managing GPU infrastructure. The service covers image generation, speech, music, image restoration, video from images, image captioning, and large language models, and it exposes all of it through a short code snippet in Node, Python, or HTTP.

A One-Line API for Open-Source Models

The core promise is simple: run and fine-tune models, deploy custom models, all with one line of code. A typical call initializes the Replicate client with an API token, passes a model and an input object, and logs the output. That pattern works across the catalog, so switching from an image model to a language model does not require a new SDK or a different deployment pipeline. The platform also offers a Playground for comparing models side by side before committing to one.

What You Can Generate and Process

Replicate's catalog spans several task families. Image generation includes models such as google/nano-banana-2, openai/gpt-image-2, google/imagen-4-ultra, black-forest-labs/flux-2-flex, and recraft-ai/recraft-v4-styles-pro-svg. Beyond still images, the platform supports speech generation, music generation, image restoration, video generation from images, image captioning, and large language models. Run counts on individual models give a rough signal of adoption; for example, openai/gpt-image-2 shows 30.9M runs and google/nano-banana-2 shows 20.6M runs.

Fine-Tuning and Custom Model Deployment

Replicate is not limited to public models. Teams can deploy their own custom models using Cog, the company's open-source tool for packaging machine learning models. Private models generally run on dedicated hardware, so they do not share a queue with other users. The trade-off is billing: you pay for all the time instances are online, including setup time, idle time waiting for requests, and active processing time. Fast booting fine-tunes are the exception, billed only for the time the model is active and processing requests. Traffic spikes are handled by automatic scaling up and down.

Built for Teams That Are Not ML Experts

Replicate positions itself around speed of delivery rather than model research. The company says a team can deploy an AI feature in a day and scale to millions of users without being machine learning experts. That framing fits product teams that want to ship an AI-powered feature, such as an AI design co-pilot or an AI logo generator, without hiring a dedicated ML platform group. The API surface stays the same whether you are calling a hosted model or your own fine-tune.

Hardware and Cost Structure

Pricing is tied to the hardware a model runs on. Replicate publishes per-second and per-hour rates for CPU and GPU options, including Nvidia T4, L40S, A100, H100, and H200, with multi-GPU configurations listed as well. Additional multi-GPU A100, H100, H200, and L40S capacity is available through committed spend contracts. Because private models are billed for setup, idle, and active time, cost planning depends heavily on how often a model is actually invoked. Fast booting fine-tunes avoid idle charges, which makes them a better fit for intermittent workloads.

Where Replicate Fits

Replicate works best for developers who want API access to a wide range of open-source models and the option to deploy their own. It is less suited to teams that need a fully managed, opinionated application layer, since the platform provides infrastructure and model access rather than end-user features. The combination of a broad catalog, Cog-based packaging, and hardware-level pricing gives teams a way to prototype quickly and then move to dedicated capacity as usage grows.

Core features

  • Run open-source models for image generation, speech, music, image restoration, video from images, image captioning, and large language models
  • Fine-tune models and deploy custom models packaged with Cog, Replicate's open-source model packaging tool
  • Call models from Node, Python, or HTTP with a short code snippet
  • Compare models in the Playground
  • Deploy private models on dedicated hardware, with automatic scaling up and down based on traffic
  • Fast booting fine-tunes are billed only for active processing time

Use cases

  1. 1Generate images from text prompts using models such as google/nano-banana-2, openai/gpt-image-2, and black-forest-labs/flux-2-flex
  2. 2Generate speech, music, or video from images
  3. 3Restore images or caption images
  4. 4Run large language models through an API
  5. 5Fine-tune a model on custom data and deploy it as a private model
  6. 6Package a custom model with Cog and serve it on Replicate

Strengths and things to check

Strengths

  • One-line code samples for Node, Python, and HTTP lower the barrier to running models
  • Access to a broad catalog of open-source models across image, audio, video, and language tasks
  • Custom models can be packaged with Cog and deployed as private models
  • Private models run on dedicated hardware and scale automatically with traffic

Things to check

  • Private models are billed for setup, idle, and active time, so idle instances still cost money
  • Fast booting fine-tunes are the exception and are billed only for active processing time
  • Multi-GPU A100, H100, H200, and L40S capacity requires committed spend contracts
  • Pricing is hardware-based and varies by GPU type, so costs depend on the model and hardware selected

Pricing

Replicate lists hardware-based pricing. CPU (Small) is $0.000025/sec ($0.09/hr); CPU is $0.000100/sec ($0.36/hr); Nvidia T4 is $0.000225/sec ($0.81/hr); Nvidia L40S is $0.000975/sec ($3.51/hr); Nvidia A100 (80GB) is $0.001400/sec ($5.04/hr); Nvidia H100 is $0.001525/sec ($5.49/hr); Nvidia H200 is $0.001525/sec ($5.49/hr). Multi-GPU configurations are also listed, with additional A100, H100, H200, and L40S capacity available through committed spend contracts. Private models are billed for setup, idle, and active time, except fast booting fine-tunes, which are billed only for active processing time.

Basic information

Added
Oct 8, 2026
Platforms
Web
Chinese & English
Not marked
API
Not marked
Self-hosting
Not marked
Best for
Developers, product teams, and startups that need to run or fine-tune machine learning models through an API without managing their own GPU infrastructure.
Information updated
Oct 16, 2026
Website status
Checking
Corrections · Privacy · Data requests?
Contact us

Features, pricing and availability may change. Always verify details on the official website. Maintained by steven · Last updated Oct 16, 2026 →

Related tools

All categories

LlamaIndex

Enterprise Grade

LlamaIndex builds document parsing and extraction software plus open-source tools for AI applications, helping teams turn enterprise documents into usable data for AI agents and workflows.

Free TierSOC 2
View details