MUGGS OF DREAMS

AI doesn't have to live in a warehouse. Some AI models are released openly: you can download their files and run them on an ordinary, if powerful, home computer. That's called running a model locally.

MuggsOfDreams runs most of its AI locally. The pages are built at home and the art is made at home, with cloud AI helping for now.

The finished pages are hosted on Cloudflare (card 05), ordinary web hosting, not AI. Being 100% data-center free is the goal by launch day.

Where cloud AI still fits, and how it's leaving, is explained below in what is ai-data center free?.

A desktop in my house with an NVIDIA RTX 4090. On it runs Qwen 3.8 27B, an open model from Alibaba's Qwen team. The files are free to download, and it all runs on this machine.

It writes and edits code. It reads about 128,000 tokens at once, and a token is a piece of a word, so that's about 96,000 words, enough for a whole page plus its notes and the instructions. It looks at pictures, so it checks its own screenshots. And it thinks a problem through before it answers.

I chose it because it's strong at code for its size, it sees images, and it fits on one home graphics card, about 17 GB once compressed.

Its job: OpenCode (card 04) drives it. Qwen builds and edits every page from a written brief, runs the house checks, and ships. The checks are my words, every fact from the owner, the layout rules, even the spacing.

By launch it'll read a neighbor's form directly and offer a few options for each part, so I pick the one that fits.

A second desktop in my house, with an NVIDIA RTX 5070 Ti, joined to the builder by a cable straight between the two. Its one job is making images. It runs two open image models from Black Forest Labs.

FLUX.2 klein 4B redraws from pictures. It takes one of my photos plus a description and paints it new. That's how a photo of the block becomes a comic-book scene, the street, the telephone pole, the flyer on it. About 5 to 7 seconds a picture.

FLUX.1 schnell draws from words alone. Schnell is German for fast, four steps a picture.

Neither can be trusted to spell, so words in a scene, a sign or a name, are set afterward in the page's own Carey font.

I chose them for three reasons. The license, Apache 2.0, allows commercial use. They fit on the 5070 Ti. And they're fast enough to make a sheet of tries.

Its job: each page's doubloon and the comic-book scenes of the neighborhood. I pick every image myself, from the numbered sheet. One firm rule: art on a live page only comes from models whose licenses allow commercial use.

A model on its own just answers questions. To build something, it needs a workbench: a place to open files, write code, test it, and keep track of what's done. That's OpenCode, a free, open-source coding agent. It isn't locked to one AI company, so you can point it at any model, including one running on your own machine.

Here, OpenCode talks to Qwen (card 02). The bridge is LM Studio, a free app that downloads open models and runs them on your own computer. The open-source engine inside it is llama.cpp, the part that runs the model on an ordinary computer, including a home graphics card. LM Studio also hands the running model to other programs on the same machine, and that's how OpenCode reaches Qwen.

ComfyUI is open-source software for running image models. On the art shop (card 03), it's the program that runs them.

Before a page ships, Python scripts check it with Playwright, an open-source tool that opens the page in a browser and measures it.

The code lives on GitHub, which keeps it and every change ever made to it. There are two copies. One is private and has everything, notes, tools, briefs. The other has only the finished site.

Cloudflare puts that site-only copy on the web. muggsofdreams.net is published from it, and each yourname.muggsof.net page is served by a small Cloudflare program that also sends the contact emails ([email protected]).

This is ordinary web hosting, not AI. No model runs there. It's what keeps the pages fast for everyone.

I'm Sean, a computer science teacher in New Orleans. The machines do the typing, but I'm responsible for what goes live.

Every page is checked against the house rules before it ships. By launch, I'll be reading every form myself, choosing among the options the model offers, and keeping a record of those decisions. When something needs a human touch, it gets one. I'm not a graphic designer, and I'll tell you when it matters.