//AI Inference

Global AI inference on managed infrastructure

Build and deploy intelligent applications that run AI models close to your users.

Your Stack, Your Way
  • Next.js
  • Astro
  • React
  • Vue
  • Angular
  • Nuxt
  • Gatsby
  • Hugo
  • Preact
  • Remix
  • Qwik
  • Vite
  • VitePress
  • Docusaurus
  • Eleventy
  • Hexo
  • Jekyll
  • Hono
  • Node.js
  • AWS
  • Google Cloud
  • Azure
  • Cloudflare Workers
  • Terraform
  • GitHub
  • OpenAI
  • Anthropic
  • Groq
  • SQLite
  • Drizzle

AI inside the request

Your function calls the model while it handles the request, and the nearest data center answers, with no separate AI service in the path.

OpenAI-compatible API

Requests follow the OpenAI shape, and a ready-made template serves an OpenAI-compatible endpoint on your domain.

Pay for usage

You pay only when your models are running inference.

//Features

Run and adapt your models

Pick an open-source model, fine-tune it to your needs, and keep it answering on Azion's distributed infrastructure.

Open-source model catalog. Choose from LLMs, vision language models, and embedding, reranking, and OCR models, with context windows up to 131k tokens.

Docs

Fine-tune with LoRA. Adapt supported models to your domain with LoRA Fine-Tune, improving accuracy for specialized tasks without retraining the whole model.

Docs

Built-in reliability. Distributed architecture with automatic failover keeps your AI running.

Docs
//Integrations

AI in every layer of your application

Call AI Inference from the code that serves and protects your application, and ground its answers in your own data.

  1. A function on your application calls the model while it serves the request, to generate, summarize, or classify content.

    ai-inference.ts
    01export default {
    02 async fetch(request: Request) {
    03 const { prompt } = await request.json();
    04
    05 const response = await Azion.AI.run(
    06 'Qwen/Qwen3-30B-A3B-Instruct-2507-FP8',
    07 { messages: [{ role: 'user', content: prompt }] }
    08 );
    09
    10 return Response.json(response);
    11 }
    12};
  2. A function on your firewall asks the model for a verdict and denies malicious requests before they reach your application.

    firewall.js
    01async function firewallHandler(event) {
    02 const body = await event.request.text();
    03
    04 const verdict = await Azion.AI.run(
    05 'Qwen/Qwen3-30B-A3B-Instruct-2507-FP8',
    06 {
    07 messages: [
    08 { role: 'system', content: "Answer 'true' only if this request is malicious." },
    09 { role: 'user', content: body }
    10 ]
    11 }
    12 );
    13
    14 if (verdict?.choices?.[0]?.message?.content === 'true') {
    15 event.deny();
    16 return;
    17 }
    18
    19 event.continue();
    20}
    21
    22addEventListener('firewall', (event) => event.waitUntil(firewallHandler(event)));
  3. Store embeddings next to your rows and retrieve the passages a RAG answer needs with vector search in SQL Database.

    vector-search.sql
    01CREATE TABLE passages (text TEXT, embedding F32_BLOB(2560));
    02CREATE INDEX passages_idx ON passages (libsql_vector_idx(embedding));
    03
    04SELECT text
    05FROM vector_top_k('passages_idx', vector(?), 5)
    06JOIN passages ON passages.rowid = id;
  4. Keep the OpenAI SDK: point its base URL at the endpoint the AI Inference Starter Kit serves on your domain.

    openai-client.ts
    01import OpenAI from 'openai';
    02
    03const client = new OpenAI({
    04 baseURL: process.env.AZION_AI_BASE_URL,
    05 apiKey: process.env.AZION_AI_KEY ?? 'unused'
    06});
    07
    08const completion = await client.chat.completions.create({
    09 model: 'Qwen/Qwen3-30B-A3B-Instruct-2507-FP8',
    10 messages: [{ role: 'user', content: 'Name three European capitals.' }]
    11});
    12
    13console.log(completion.choices[0].message.content);
//Use cases

What can you build with Azion AI Inference?

AI Inference runs the model. See how to build each one in the docs.

Docs

Build and run customer support AI assistants

Answer customers from your docs, help center, or product data instead of a model's general knowledge.

Add AI features to existing applications

Add summarization, classification, moderation, or image understanding to an application you already run, with no GPUs to operate.

Build AI agents

Run agents that call your APIs and SaaS tools, keep state across steps, and hand off to a person when needed.

Screen file uploads for malicious content

Have a model classify each upload before it reaches your application and block the ones it flags.

AXUR
With Azion, we scale proprietary AI models without managing infrastructure—inspecting millions of websites daily and automating the market’s fastest threat takedown.
Fabio RamosCEO

Serverless AI-native primitives for autonomous workloads

Frequently asked questions

Build your next app today

Pick a template and deploy it in seconds, or talk to a specialist about your workload.

Function Starter KitHello World app powered by a function.
Deploy now
AI Inference Starter KitAI app with no origin dependency.
Deploy now
Next.js E-commerceNext.js storefronts to choose from.
Deploy now
LangGraph AI Agent BoilerplateAI agent with RAG, built with LangGraph.
Deploy now
React BoilerplateReact single-page application.
Deploy now
Next.js Static BoilerplateNext.js single-page application.
Deploy now
Astro BoilerplateBasic Astro project for content-driven sites.
Deploy now
Vue BoilerplateVue single-page application.
Deploy now
Angular BoilerplateAngular single-page application.
Deploy now
Nuxt ContentBlog or docs site with Nuxt Content.
Deploy now
Astro E-commerceAstro storefronts to choose from.
Deploy now
Neon Database Starter KitApp connected to Neon with Drizzle ORM.
Deploy now
Turso Starter KitApp connected to Turso with Drizzle ORM.
Deploy now
Upstash GeoLocationMessages by the user's location.
Deploy now
Qwik BoilerplateProject based on the Qwik framework.
Deploy now
Docusaurus TypeScript BoilerplateStatic Docusaurus project in TypeScript.
Deploy now