Run AI
with an API.

Run and fine-tune open-source models. Deploy custom models at scale. All with one line of code.

With Replicate you can

Thousands of models contributed by our community

All the latest open-source models are on Replicate. They’re not just demos — they all actually work and have production-ready APIs.

AI shouldn’t be locked up inside academic papers and demos. Make it real by pushing it to Replicate.

How it works

You can get started with any open-source model with just one line of code. But as you do more complex things, you fine-tune models or deploy your own custom code.

Run open-source models

Our community has already published thousands of models that are ready to use in production. You can run these with one line of code.

import replicate

output = replicate.run(
  "stability-ai/sdxl:39ed52f2a78e934b3ba6e2a89f5b1c712de7dfea535525255b1aa35c5565e08b",
  input={
    "width": 768,
    "height": 768,
    "prompt": "An astronaut riding a rainbow unicorn, cinematic, dramatic",
    "refine": "expert_ensemble_refiner",
    "scheduler": "K_EULER",
  }
)

print(output)

Fine-tune models with your own data

You can improve open-source models with your own data to create new models that are better suited to specific tasks.

Image models like SDXL can generate images of a particular person, object, or style.

Language models like Llama 2 generate text in a specific style or get better at a particular task.

Train a model:

import replicate

training = replicate.trainings.create(
    version="stability-ai/sdxl:c221b2b8ef527988fb59bf24a8b97c4561f1c671f73bd389f866bfb27c061316",
    input={
        "input_images": "https://my-domain/my-input-images.zip",
    },
    destination="mattrothenberg/sdxl-fine-tuned"
)

print(training)

This will result in a new model:

mattrothenberg

mattrothenberg/sdxl-fine-tuned

A very special, fine-tuned version of SDXL

0 runs

Then, you can run it with one line of code:

output = replicate.run(
    "mattrothenberg/sdxl-fine-tuned:abcde1234...",
    input={"prompt": "a photo of TOK riding a rainbow unicorn"},
)

Deploy custom models

You aren’t limited to the models on Replicate: you can deploy your own custom models using Cog, our open-source tool for packaging machine learning models.

Cog takes care of generating an API server and deploying it on a big cluster in the cloud. We scale up and down to handle demand, and you only pay for the compute that you use.

First, define the environment your model runs in with cog.yaml:

build:
  gpu: true
  system_packages:
    - "libgl1-mesa-glx"
    - "libglib2.0-0"
  python_version: "3.10"
  python_packages:
    - "torch==1.13.1"
predict: "predict.py:Predictor"

Next, define how predictions are run on your model with predict.py:

from cog import BasePredictor, Input, Path
import torch

class Predictor(BasePredictor):
    def setup(self):
        """Load the model into memory to make running multiple predictions efficient"""
        self.model = torch.load("./weights.pth")

    # The arguments and types the model takes as input
    def predict(self,
          image: Path = Input(description="Grayscale input image")
    ) -> Path:
        """Run a single prediction on the model"""
        processed_image = preprocess(image)
        output = self.model(processed_image)
        return postprocess(output)

Scale on Replicate

Thousands of businesses are building their AI products on Replicate. Your team can deploy an AI feature in a day and scale to millions of users, without having to be machine learning experts.

Automatic scale

If you get a ton of traffic, Replicate scales up automatically to handle the demand. If you don't get any traffic, we scale down to zero and don't charge you a thing.

  • CPU $0.000100/sec
  • Nvidia T4 GPU $0.000225/sec
  • Nvidia A40 GPU $0.000575/sec
  • Nvidia A40 (Large) GPU $0.000725/sec
  • Nvidia A100 (40GB) GPU $0.001150/sec
  • Nvidia A100 (80GB) GPU $0.001400/sec
  • 8x Nvidia A40 (Large) GPU $0.005800/sec
  • Learn more about pricing

Pay for what you use

Replicate only bills you for how long your code is running. You don't pay for expensive GPUs when you're not using them.

Abstract squares illustration

Forget about infrastructure

Deploying machine learning models at scale is hard. If you've tried, you know. API servers, weird dependencies, enormous model weights, CUDA, GPUs, batching.

Prediction throughput (requests per second)

Logging & monitoring

Metrics let you keep an eye on how your models are performing, and logs let you zoom in on particular predictions to debug how your model is behaving.