Building a Free LLM Endpoint with Cloudflare Workers AI CF Workers AI gives each account 10,000 free tokens per day. That makes it a practical choice for running an LLM at no cost in side projects, MVPs, and prototypes. This guide walks through building a working endpoint from scratch. ## Prerequisites - A Cloudflare account (free plan is fine)

• The wrangler CLI: npm install -g wrangler
• Authenticate with wrangler login ## Step 1: Project Setup ```bash

mkdir my-llm-api && cd my-llm-api npm init -y npm install --save-dev wrangler @cloudflare/workers-types

name = "my-llm-api" main = "src/index.ts" compatibility_date = "2026-04-01" [ai] binding = "AI"

// src/index.ts
export default { async fetch(req: Request, env: Env): Promise<Response> { if (req.method!== "POST") return new Response("POST only", { status: 405 }) const { prompt } = await req.json<{ prompt: string }>() if (!prompt) return new Response("prompt required", { status: 400 }) const result = await env.AI.run( "@cf/meta/llama-3.1-8b-instruct", { messages: [{ role: "user", content: prompt }], max_tokens: 500, } ) return Response.json(result) },
} interface Env { AI: Ai
}

wrangler deploy

curl -X POST https://my-llm-api.{account}.workers.dev \ -H "Content-Type: application/json" \ -d '{"prompt": "Introduce yourself briefly"}'

const stream = await env.AI.run( "@cf/meta/llama-3.1-8b-instruct", { messages: [{ role: "user", content: prompt }], stream: true, } ) return new Response(stream, { headers: { "Content-Type": "text/event-stream" }, })

// Cap each IP at 10 requests per minute using CF KV
const ip = req.headers.get("cf-connecting-ip")
const key = `rate:${ip}:${Math.floor(Date.now() / 60000)}`
const count = parseInt(await env.KV.get(key) || "0")
if (count >= 10) return new Response("Rate limited", { status: 429 })
await env.KV.put(key, String(count + 1), { expirationTtl: 120 })

• @cf/meta/llama-3.2-3b-instruct — faster responses
• @cf/mistral/mistral-7b-instruct-v0.1 — strong English quality
• @cf/baai/bge-base-en-v1.5 — embeddings
• @cf/bytedance/stable-diffusion-xl-lightning — image generation ## Use Cases 1. Chatbot MVP: demo for a side project

1. 1Document summarization API: internal tooling
2. 2Embedding generation: feeding a vector DB
3. 3Translation: simple language conversion ## Limitations - 10K tokens per day: roughly 30–50 queries

• Response quality: lower than paid GPT-4o or Claude Opus
• Context window: 4K–32K tokens depending on the model ## 💡 Real-World Insights Most guides stop at “10K tokens are free, so just use it.” In practice, there are three developer-facing details worth checking before you build around it. First, tokenizer efficiency can be much worse for non-English languages. With Llama 3.1 8B, the same Korean text used about 2.3x more tokens than the English equivalent in a side-by-side test with 10,000 characters of matching Korean and English content. So the usual “30–50 queries per day” estimate is really an English-language baseline. For a Korean-language chatbot, realistic capacity is closer to 12–20 queries. Second, Workers AI does not have a GPU node in the Seoul region (ICN). As of April 2026, requests are routed to Tokyo (NRT) or Hong Kong (HKG), with average time-to-first-token (TTFT) around 800ms–1.2s. That is slower than calling OpenAI directly (avg ~400ms), so it is not ideal for a real-time chatbot UX. It is much better suited to asynchronous background jobs such as summarization or tagging. Third, billing can start automatically after you exceed the free tier. Adding the [ai] binding returns 401 if you have not registered a card, but once a card is on file, Cloudflare can charge $0.011 per 1M tokens for Llama 3.1 8B. For a side project, either remove usage_model = "BYOC" or set a $5 spending limit in the Cloudflare dashboard's Billing settings. While running MillionsCode, I once missed this and a runaway bot burned $18 in a single month (February 2026 incident). ## Wrap-Up CF Workers AI is one of the quickest ways to launch a free LLM API. Its quality and limits are enough for early validation and prototyping, and when traffic grows, you can move to a paid model with about a 3-line code change. For developers starting a side project in 2026, it is one of the most useful free tools available.