Free Hosting for AI Apps in 2026: Vercel, Cloudflare, Hugging Face Spaces, Render, Fly.io, Cloud Run

An AI app is usually a thin web app: a page, a server route that holds your API key, and calls to a model hosted by someone else. That needs little compute, so a free host can carry it, as long as you know what the free plan stops you from doing. Free plans differ most in three places: whether the app may earn money, whether it sleeps when idle, and how long one request may run (LLM calls are slow). Everything below was read on each vendor's pricing or docs page on 2026-10-28.

Free hosting limits side by side

Source: each vendor's pricing or docs page, linked in the table and under Sources, checked 2026-10-28.

Hosts that fit an AI app

• Vercel Hobby is the easiest for a Next.js chat app: a function can run up to 300 seconds, enough for a slow LLM answer. The rule that matters is on its fair use page: the Hobby plan "restricts users to non-commercial, personal use only". A side project is fine; a paid product is not. Pro is a paid plan. See the offer page for Vercel's startup program if you qualify.
• Cloudflare Workers gives 100,000 requests a day with no sleeping, but only 10 ms of CPU time per invocation. Time spent waiting for a model API is not CPU time, so proxying an LLM call is usually fine; heavy parsing in the Worker is not. Details: Workers free plan.
• Workers AI runs open models on Cloudflare's network for 10,000 Neurons a day. How many answers that is depends on the model; a small model gives many more than a 70B one. See Workers AI free tier.
• Deno Deploy and Netlify are the other no-sleep options; see the Deno Deploy plan and the Netlify free plan.

Hosts with a catch

• Hugging Face Spaces used to be the default place to put a Gradio demo for free. On 2026-10-28 its docs say: "Gradio and Docker Spaces run on compute and require a paid plan to create". Static Spaces are free, and a free account can still run up to 2 Gradio Spaces on ZeroGPU. See ZeroGPU Spaces.
• Render is a real server for free, but it sleeps: after 15 minutes without traffic the service spins down and the next visitor waits about a minute. The free Postgres database is deleted 30 days after creation (14 days of grace to upgrade). Fine for a demo you show someone, bad for a public app.
• Fly.io has no free tier for new organizations. Accounts on plans discontinued in October 2024 keep their old allowances. Anyone else gets a short trial and then adds a card.
• Google Cloud Run has a real monthly free amount and scales to zero, but it sits behind a billing account, so a bug that loops can cost money. Set a budget alert. Google's other free products are in the Google Cloud free program.
• Railway is closer to a trial than a free plan: $1 of usage a month after the $5 trial credit. See Railway.

Which one to pick

• A hobby chat app in Next.js: Vercel Hobby. Keep it non-commercial.
• An app that might earn money: Cloudflare Workers or Deno Deploy, then check their terms; avoid Hobby.
• A demo with a model on a free GPU: a ZeroGPU Space on Hugging Face.
• A Python server that must stay up between requests: Render sleeps (about a minute to wake); Cloud Run scales to zero but needs a billing account.
• A static front end with a separate API: GitHub Pages or Cloudflare static assets.

Pair the host with a model: the chat app stack and the tool-calling agent stack pick a free LLM API and a free host from the catalog, and the RAG stack adds a vector database.

Steps

1. Put the app in a Git repository and keep the model API key in an environment variable, never in the browser code.
2. Connect the repository to the host (Vercel, Netlify, Render and Cloudflare all deploy on push).
3. Add the API key as a secret in the host's settings and deploy.
4. Open the live URL, send three prompts and look at the host's usage page to see what they cost in requests or CPU time.
5. Add a spending alert where a card is on file (Cloud Run, Fly.io).

Limits and the catch

• Free plans change. Fly.io removed its free allowance in 2024 and Hugging Face now asks for a paid plan to create Gradio Spaces; re-read the pricing page before you rely on a limit.
• Commercial use differs. Vercel Hobby and GitHub Pages forbid running a business on them; check the terms before you take a payment.
• Sleeping hosts are slow for the first visitor. Render's wake-up takes about a minute.
• Request time matters for LLMs. A model answer can take many seconds: check the maximum duration of a function or request on the plan you choose.
• Your visitors spend your API quota. A free host does not cap the model bill or rate limit; see rate limits.