Replicate
Run and deploy any open ML model — LLMs, image, video, audio — through one API with pay-per-second billing, and package your own with open-source Cog.
Replicate is a hosted platform for running open machine-learning models — LLMs, image, video, audio, and more — behind a single API, billed per second of compute. Thousands of community models run with one line of code, and you can package and deploy your own using Cog, its open-source container tool. Replicate joined Cloudflare in 2025.
Replicate is a hosted platform for running and deploying open machine-learning models through an API. Instead of provisioning GPUs, building containers, and writing inference servers, you call a model — Stable Diffusion-style image generators, video and audio models, open LLMs, and thousands of other community-contributed models — with a single line of code and get results back. The platform handles the GPUs, scaling, and queuing.
Its defining design choice is pay-per-second compute: on public models you are billed only for the time your code actually runs, which makes it well suited to spiky, experiment-heavy, or low-volume workloads where a dedicated endpoint would mostly burn money. Private models work differently: most run on dedicated hardware and are billed for all the time their instances are online, including setup and idle time, except fast-booting fine-tunes. Some models are billed per output (per image, for example) rather than per second. New accounts can try a curated collection of models free for a limited number of runs before moving to prepaid credit.
The bring-your-own-model path runs through Cog, Replicate's open-source (Apache-2.0) container tool. You describe a model's environment and inputs in a single config file, and Cog produces a reproducible container with an HTTP inference API and CUDA/dependency setup handled for you — then you push it to Replicate to get an autoscaling endpoint, identical in shape to the community models. Against infrastructure platforms like Modal and Baseten, Replicate leans hardest on breadth of ready-to-run open models and the one-line API, where the others center custom code and deployment control.
Replicate joined Cloudflare in 2025: Cloudflare announced the acquisition in November 2025 and closed it on December 1, 2025, and Replicate said it would carry on as a distinct brand with the same API. In April 2026 Cloudflare said the Replicate team had joined its AI Platform team and that it is moving Replicate's models onto Cloudflare infrastructure and into AI Gateway. For weighing a hosted inference API against running models yourself, see self-host vs. API LLM.
Frequently asked questions
- What is Replicate?
- Replicate is a cloud platform that runs and deploys open machine-learning models behind a single HTTP API. It hosts thousands of community-contributed models — LLMs, image, video, audio, and more — that you call with one line of code, no GPU provisioning or container plumbing required. You can also push your own model with Cog and serve it the same way.
- How does Replicate pricing work?
- Pay-per-use, not a flat subscription. Most models are billed by the second of compute, while some (e.g. certain image models) are billed per output. New accounts run a curated set of models free for a limited number of runs, then move to prepaid credit. Public models are billed only for the time they actually run, but most private models run on dedicated hardware and are billed for all the time their instances are online, including setup and idle time; fast-booting fine-tunes are the exception.
- What is Cog and how do I deploy a custom model?
- Cog is Replicate's open-source (Apache-2.0) tool for packaging an ML model into a standard, production-ready container — it generates the inference API and handles CUDA/dependency setup from a single config file. You define your model with Cog, push it to Replicate, and it becomes a deployable API endpoint with autoscaling, the same as any community model.
Filed under
inference · open-models · gpu · image-generation · api
Related
- Self-Host vs API: When Does Running Your Own LLM Actually Pay Off?The real economics of self-hosting an LLM vs. calling a hosted API — GPU utilization, privacy, latency, and the hidden ops costs that decide the crossover.
- Deploying LLMs to Production: A Reliability & Cost ChecklistTake an LLM feature from prototype to production: API vs self-host, provider fallback, retries, caching, observability, eval gates, and safe rollout.
- Multimodal RAG over PDFs, Scans & Charts: Two Approaches That Actually WorkRAG over visual documents — PDFs, scans, charts — where text-only extraction loses tables and layout. Parse-then-text vs embed-the-page-image, with trade-offs.
- Cold Start OptimizerCut cold-start latency for serverless functions and slow-booting apps by measuring the init breakdown, then attacking the dominant phase — artifact size, eager imports, eager connections, or under-provisioned memory — instead of reflexively buying provisioned concurrency. Use when serverless p99 spikes on the first request, when a function times out during init, or when scale-to-zero is hurting user-facing latency.
- The Best LLM Inference Providers in 2026: Groq vs Together vs Fireworks vs Baseten and MoreGroq, Cerebras, Together AI, Fireworks AI, Baseten, Modal, fal and Replicate compared by workload, with 2026 ownership changes and OpenAI compatibility.
- Diffusion ModelA diffusion model generates an image by starting from random noise and removing it step by step, guided by a prompt, until a coherent picture remains.