Skip to content

About

The Local AI Purity Test: how local are you? A test for people who run open-weight models, with ranks from Curious to Legend.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Local AI Test logo

The Local AI Purity Test

A test for people who run open-weight models. Tick everything you've done and your score earns you a rank from Curious to Legend.

Take it at localaitest.com

The Local AI Purity Test: How local are you?

How it works

There are 100 items. Your score is how many you ticked, so higher means more local.

Score Rank
0–5 Curious
6–15 Enthusiast
16–30 Explorer
31–45 Regular
46–60 Builder
61–75 Quantizer
76–90 Power User
91–99 Legend
100 ???

Every rank has its own animated card with a roast and a score of exactly 0 gets one too. Every score has its own share link (like /r/52-rich) and its link preview shows the rank's card.

A rank card: You are Builder.

Share on 𝕏 copies the card and opens a post or opens the share sheet on phones. Download video saves the animated card to attach instead.

Rig badge

Optionally add your hardware, either read from the hardware listed on your Hugging Face profile or picked by hand. Your total GPU or unified memory earns a badge:

Total memory Badge
Under 16GB GPU poor
16–47GB VRAM middle class
48–191GB GPU rich
192GB and up Datacenter in a closet

Autofill

Connect GitHub or Hugging Face and the test ticks the items it can prove from your public profile. Each one is tagged "detected via…", so you can untick it.

  • Hugging Face: your listed hardware (two 3090s, a 64GB or 512GB Mac, a DGX Spark or Strix Halo, two or more Sparks, four or more Sparks), uploaded models, GGUFs, LoRAs, merges and whether your downloads beat your followers
  • GitHub: starring llama.cpp and opening issues or getting PRs merged in llama.cpp, Ollama or vLLM

The jokes

  • A VRAM bar that fills as you tick items and overflows into "offloading to CPU… 0.4 tok/s"
  • A GPU fan and temperature ("Your GPU: 68°C") that climb with your answers
  • An "Ask an AI Assistant" button that runs out of CUDA memory
  • Jokes on 37 items, shown when you hover over or tick them, including a deliberately bad pelican on a bicycle
  • A hidden phrase: type gguf when anywhere on the page
  • A 404 page that tells you to ollama pull it

Privacy

The site is run by Eigen Labs, Inc. The footer's privacy note is the user-facing version of this section; keep the two in sync.

  • Your answers never leave your browser. They're saved on your device only (localStorage) and Reset deletes them.
  • No accounts, no analytics and no tracking. The site's own code doesn't keep or log anything.
  • The hosting provider (Vercel) receives standard request data, like IP addresses, to serve the page.
  • Autofill is optional. The browser reads your public profile straight from GitHub's and Hugging Face's APIs.
    • Hugging Face login runs entirely in the browser, with PKCE. Its token can only read your basic profile and isn't cancelled afterward.
    • GitHub login asks for no permissions, so its token can only read public data. One server route swaps the login code for a token and hands it to the browser in the URL fragment, without storing or logging it. The browser asks GitHub to cancel the token when autofill finishes; if the page closes first, it isn't cancelled.
  • Share links carry only your score and rig badge.

Stack

  • Next.js 16 (App Router, Cache Components) with TypeScript and Tailwind CSS 4
  • Rank cards are pre-made images and videos in public/cards and all 505 result pages are prerendered at build time
  • The homepage preview card, favicon and logo are drawn with next/og
  • Hosted on Vercel

Project layout

Path What's there
src/data/items.ts The 100 items, their jokes and which ones can be autofilled
src/lib/score.ts Ranks, roasts, rig badges, result codes and share text
src/lib/rig.ts Hugging Face hardware parsing, the device picker list and hardware rules
src/lib/detect.ts Browser-side Hugging Face and GitHub lookups
src/lib/oauth.ts, src/app/api/auth/github/ The Connect GitHub and Connect Hugging Face logins
src/lib/og.tsx, src/lib/brand-icon.tsx The homepage preview card, favicon and logo
src/components/ The checklist, results page, status bar, autofill panel and popups
public/cards/ Each rank's card, as a video and a still frame

Develop

npm install
npm run dev

Open http://localhost:3000.

npm test        # rank, rig and share-code unit tests
npm run build   # production build; prerenders all 505 result pages

Set NEXT_PUBLIC_SITE_URL to override the share-link domain (defaults to https://localaitest.com).

The Connect buttons need these environment variables, in .env.local locally and in your host's settings for production. Without them the buttons say they "aren't set up yet" and the rest of the test still works.

Variable From
GITHUB_CLIENT_ID A GitHub OAuth app whose callback is <site>/api/auth/github/callback
GITHUB_CLIENT_SECRET The same GitHub OAuth app
NEXT_PUBLIC_HF_CLIENT_ID A Hugging Face OAuth app with no secret, scopes openid profile and redirect <site>/

Credits

Made by Pratik. Design and rank cards by Estella Xian.

Questions and ideas from Ivan Fioravanti, Kydo, @Spangler3000 and David Tai.

Inspired by the work of Tyler and Sarah on sfpurity.com.

License

The code is released under the MIT License.

The "Local AI Purity Test" name, the localaitest.com domain, the logo and the rank card artwork in public/cards are not covered by the license. If you fork this, please give it your own name and look.

The fonts are licensed separately:

  • Geist Mono is under the SIL Open Font License 1.1.
  • Zalando Sans SemiExpanded is under the SIL Open Font License 1.1. The fixed-weight copies in assets/fonts, used for generated images, are renamed "LAT Sans SemiExpanded" as that license requires for modified versions.
  • ABC Repro is a commercial font from Dinamo, licensed for localaitest.com. It isn't covered by the MIT License, so swap in your own font if you fork this.

About

The Local AI Purity Test: how local are you? A test for people who run open-weight models, with ranks from Curious to Legend.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages