Replicate

Replicate is a hosted API for running open-source machine learning models without provisioning GPUs. Developers call a community model in a line of code, fine-tune it on their own data, or package a proprietary model with the open-source Cog container format and deploy it on Replicate's infrastructure, paying per second of compute rather than for idle capacity.

Replicate is a cloud API that runs open-source machine learning models so developers never have to provision a GPU, build a serving stack or write a Dockerfile for CUDA. A prediction is a single HTTP call or a few lines of Python or Node, and the platform handles scheduling, scaling to zero and returning the output, billed per second of the hardware actually used. Underneath sits Cog, Replicate's open-source container format for packaging models, which is what lets the same interface serve a community image model, a customer's private fine-tune and a large language model without bespoke integration work in each case. Cloudflare's November 2025 acquisition announcement put the catalogue at more than 50,000 production-ready models, spanning image and video generation, speech and music synthesis, image restoration, captioning and text models. Around that core the product adds fine-tuning — training a custom FLUX or SDXL variant on a user's own images — plus Deployments for pinning dedicated capacity to a model, private models, webhooks for asynchronous jobs, organisation accounts, LoRA support and documented rate limits and data-retention policies. Pricing is published in full and unusually granular for the category: CPU from $0.000025 per second, an Nvidia T4 at $0.000225 per second, an L40S at $0.000975, an A100 80GB at $0.0014 and an H100 at $0.001525, with multi-GPU configurations scaling proportionally and some models billed per token or per generated image instead. Companies including Labelbox, Character.ai, Photo.ai, HeadshotPro and Magnific are named as building on the platform. Cloudflare announced the acquisition on 17 November 2025, with the catalogue to be folded into Workers AI.

Category

AI Models & APIs

Pricing

Usage-based, Contact for pricing

Target Market

CTOs, Enterprise Developers, Data Scientists, ML Engineers, Startup Founders

Deployment

Cloud-only, API-based

• ✓
  50,000+ model catalogue

  A community library spanning image, video, audio, language and restoration models, all callable through one consistent prediction API.

• ✓
  Cog packaging format

  An open-source container standard for ML models that makes a custom model deployable with the same interface as any community model.

• ✓
  Per-second hardware billing

  Published rates from $0.000025/sec on CPU to $0.001525/sec on an H100, so idle time between requests costs nothing.

• ✓
  Fine-tuning

  Train a custom variant of models such as FLUX or SDXL on your own dataset and serve it through the same endpoint.

• ✓
  Deployments

  Pin dedicated capacity to a specific model version to control scaling behaviour and reduce cold-start latency for production traffic.

• ✓
  Webhooks and async predictions

  Long-running generations report back by webhook rather than holding an HTTP connection, which is essential for video and large image jobs.

• ✓
  Private models and organisations

  Keep proprietary weights private and manage access, API tokens and billing across a team from one account.

• •
  Adding image generation to a SaaS product

  Call a hosted diffusion model from application code and ship a generation feature without operating any GPU infrastructure.

• •
  Custom fine-tuned avatars and headshots

  Train a per-customer model on uploaded photos, then serve each user's private fine-tune through the same prediction API.

• •
  Model evaluation and bake-offs

  Run the same prompt across a dozen competing open-source models to pick one, paying only for the seconds consumed.

• •
  Speech, music and audio pipelines

  Chain transcription, synthesis and enhancement models into a pipeline without managing separate serving stacks for each.

• •
  Deploying a proprietary in-house model

  Package a private model with Cog and get autoscaling inference plus webhooks without building a serving layer.

Billing is per second of the hardware a prediction occupies, with every rate published openly — $0.000225/sec for a T4 up to $0.001525/sec for an H100 — and some models metered per input token, output token or generated image instead. Fast-booting fine-tunes are charged only for active processing, not idle time. There is no free tier; all usage is paid. Volume discounts, SLAs and a dedicated account manager sit behind an enterprise conversation. The economics favour spiky traffic: practitioners report the serverless premium runs well above on-demand GPU pricing at sustained load.

This page was written from 6 sources, 3 on domains other than replicate.com.

1. 1.replicate.com — pricingvendor
2. 2.replicate.com — replicate.comvendor
3. 3.replicate.com — docsvendor
4. 4.cloudflare.com — cloudflare to acquire replicate to build the most seamless a
5. 5.hn.algolia.com — hn.algolia.com
6. 6.techcrunch.com — replicate wants to take the pain out of running and hosting