Inference Endpoints

What it does

Inference Endpoints takes an AI model, usually one from the Hugging Face Hub, and runs it as a hosted service that accepts requests and returns model outputs. You choose a model, select hardware such as GPU or CPU instances, and pick an inference engine such as vLLM, SGLang, llama.cpp, TGI, TEI, or a custom container. The platform handles provisioning, scaling, and serving so the endpoint can be called from an application.

Teams use it to move a model from testing into production without managing servers. Capacity scales up and down with traffic, logs and metrics support debugging, and model weights download securely from the Hub. A catalog of ready-to-deploy models supports one-click deployment, and the hf CLI or MCP integration lets a coding agent create an endpoint from a single prompt.

It suits developers and teams deploying their own or Hub-hosted models who want managed infrastructure rather than running Kubernetes or CUDA themselves. You need a Hugging Face account to start. Costs are based on instance hours, so usage should be monitored, and enterprise terms require a quote.

What it can do

• Deploy Hugging Face Hub models or custom containers to dedicated endpoints
• Autoscale compute with traffic
• Built-in logs and metrics for observability
• Choose inference engines such as vLLM, SGLang, llama.cpp, TGI, and TEI
• Fast, secure weight downloads from the Hugging Face Hub
• One-click deployment from a catalog of ready-to-deploy models

Works with

• Hugging Face Hub
• hf CLI
• MCP

Example workflows

Deploy a catalog model with one click

Input: a chosen model from the catalog and a GPU instance. Steps: browse the catalog, select the model and hardware, deploy. Output: a running dedicated endpoint that can receive inference requests.

Create an endpoint from a coding agent

Input: a prompt naming a Hub model, with the hf CLI installed. Steps: send the prompt to the agent, which uses the hf CLI to configure an endpoint. Output: an optimized Inference Endpoint for that model.

Before you choose

• Costs accrue per instance hour while the endpoint runs, so pause or remove unused endpoints.
• Model support depends on the selected inference engine and hardware; verify compatibility before deploying.
• Enterprise or large-scale needs require a quote; request one from Hugging Face for terms beyond the listed rates.

Listed under

• Coding
• Sandboxes and infrastructure
• AI infrastructure
• AI models
• Cloud hosting
• Dedicated endpoints
• Deploy models
• Developers

Get started

Sign up for a Hugging Face account, then create a first Endpoint.

1. Create a Hugging Face account and sign in.
2. Open the endpoint creation flow and choose a model from the Hub or catalog.
3. Select an instance type and inference engine, then deploy and test with a request.

Charged by dedicated instance hour and selected hardware, with monthly figures listed as examples. Check https://huggingface.co/pricing#endpoints for current rates. Whether a free tier applies is not confirmed.

Source: huggingface.co/pricing#endpoints, checked 28 Oct 2026. Prices can change; check before you buy.

Sources for this overview

Product pages, documentation, and pricing checked 28 Oct 2026. Each source supports part of the overview; availability and prices can change.

• Inference Endpoints by Hugging Face ↗endpoints.huggingface.co
• Inference Endpoints · Hugging Face ↗huggingface.co
• Pricing - Hugging Face ↗huggingface.co

Company website

endpoints.huggingface.co: product, pricing and policy pages

YC directory mirror

Batch and company name

Web search

Confirms the official site

Social accounts

Accounts the company links, and its posts on X

Human review

Checked by a person

knowledge_base

Last checked 28 Oct 2026. Every field keeps its source and date; unknown means we couldn’t confirm it.