A grey Super Nintendo console with two controllers, the kind of vintage hardware ChatGPT for SNES now runs on.

ChatGPT for SNES: a 2.2M-parameter chatbot on a 1990 console

Photo: France1978 on Flickr, CC BY 2.0

Someone finally asked the question nobody at OpenAI dared to: how good would ChatGPT be if it ran on the best GPU of 1995? The answer is ChatGPT for SNES, a 2.2-million-parameter language model squeezed onto a Super Nintendo cartridge, complete with chiptune music and clicky typing sound effects. It’s useless, it’s wrong half the time, and I love it.

The man behind it is Benj Edwards, a tech journalist and retro-computing nut who has spent the last year theorizing about what he calls “tiny talking machines”. He’s now published the thing as a playable ROM, and you can try it in your browser right now via an embedded SNES9x emulator. No install, no account, no “sign in with Google”, which already puts it ahead of most AI products launched this year.

What ChatGPT for SNES actually is

Let’s be clear about the word “ChatGPT” here. The model is a distilled offshoot of GPT-OSS-120b, OpenAI’s open-weight model, which was used to generate its training conversations. So it’s a ChatGPT descendant the same way a goldfish is a descendant of whatever crawled out of the ocean. The base build weighs about 4 MB and knows roughly 8,000 words.

There’s a second version at 19 MB that bolts on a “fact book”: the opening sentences of around 150,000 Simple English Wikipedia articles, stored in an MSU-1 data file. That’s the one you want. Edwards himself admits the pure model “often gets things wrong, sometimes confidently”, and that the fact book “can carry mistakes too”, as The Register noted after poking at both builds. So, basically a regular chatbot, but honest about it.

It’s the same kind of “why not?” energy as the port that put Unreal Tournament 2004 in a browser tab last week. Nobody needs it. Everybody clicks.

The Super FX 2 chip does the heavy lifting

The clever bit is the hardware choice. The SNES main CPU is a 16-bit Ricoh 5A22 running at a heroic 2.68 MHz, which is not what you’d call an inference monster. So Edwards leaned on the Super FX 2, the RISC coprocessor Nintendo stuffed into cartridges like Yoshi’s Island and Doom to push polygons. It runs at 21.47 MHz, roughly eight times the main chip.

Edwards says he has a version that works without the Super FX at all, “but without the extra processing power the answers are not as good”. Which is a sentence I never expected to read about a Nintendo graphics chip and a language model in 2026, and yet here we are.

Want it on real hardware? You’ll need a flash cart with Super FX support, like the FXPak Pro or SD2SNES, plus MSU-1 support if you want the fact book. Think of it as the retro cousin of that $45 USB-C hub someone turned into a Switch 2 dock: cheap-ish parts, a ROM, and a weekend.

AI wrote the code, a human had the idea

Here’s the part that’ll annoy some of you: Edwards didn’t write the code. Coding agents did. He designed the scope, the chip choice, the interface, even the sounds it makes while it types, and treats the whole thing as a human/AI collaboration. Funnier still, the agents told him they’d hit the ceiling of what the hardware can do. He’s not buying it: “It’s up to human creativity to transcend those assumptions.”

I’d normally roll my eyes at “AI collaboration” talk, but this is the good version of it. No pitch deck, no “reimagining productivity”, just a guy using big models to build absurdly small ones because the idea made him grin. He’s also done an NES ROM called Séance, where an even tinier model plays a ghost trapped in a haunted cartridge. That’s more product vision than most Series B startups.

My take: go play with the 19 MB build for ten minutes. Ask it what a cat is, ask it to add 17 and 25, then ask it something about history and enjoy the confident nonsense. It’s a toy, and it’s a sharp reminder that the hardware we called obsolete decades ago still has more headroom than we think. The FAA still runs on floppy disks, after all.

Thirty-five years on, the SNES has an AI assistant. Yours probably still can’t dock a Switch properly.


Discover more from TechPassion

Subscribe to get the latest posts sent to your email.

Comments
Add a comment

Leave a Reply