3 Sources
[1]
ChatGPT shrunk down to run on a Super Nintendo
Just don't expect complex, coherent, or always correct answers If you've ever wondered how small an instance of ChatGPT could get, wonder no longer. Someone has actually rigged up a tiny version of the chatbot to run on a Super Nintendo cartridge. Tech journalist and retro tech enthusiast Benj
[2]
Someone made ChatGPT run on a real SNES, and it has 2.2M parameters
* A 2.2M-parameter ChatGPT runs on a Super NES, squeezing modern AI into retro hardware. * It answers simple questions surprisingly well but is slow and can go off the rails. * Try the interactive ChatGPT-for-SNES demo in your browser to see the quirky results. As AI models get more powerful,
[3]
Someone ran ChatGPT on an SNES, and it performs better than you'd expect
One of my favorite things to do with modern tech is to see just how far I can strip it down. While we often think of LLM models as cutting-edge technology that requires powerful hardware to use, or at least a solid Linux distro, this reporter proved that he could run it on real SNES hardware. Add
Share
Copy Link
Benj Edwards, a tech journalist and retro-computing enthusiast who runs the Vintage Computing site, has built a chatbot that runs on actual Super Nintendo hardware
3
1
. The result is a 2.2-million-parameter model that lives on a cartridge, small enough to fit where 1990s games stored their level data2
3
.
Source: How-To Geek
He announced the project on X with a question: "How good would @ChatGPT be if it ran on the Super NES's best GPU, the Super FX 2 chip?"
3
.The Super FX 2 is a 21 MHz coprocessor that debuted in 1995 with Yoshi's Island, where it handled graphics scaling and sprite rotation
3
. The chip reads the model's 8-bit weights straight off the cartridge, per How-To Geek's writeup. It timed the result at about two tokens per second, against 80 to 100 or more for the standard ChatGPT, so each word takes roughly half a second3
.The build itself leaned on bigger AI. Training conversations came from OpenAI's open-weight gpt-oss-120b, and AI coding agents wrote the code while Edwards designed what the project would do
3
1
.There are two versions
3
1
. A pure model fits in about 4MB. A factbook version, around 19MB, checks a local fact book built from the opening sentences of roughly 150,000 Simple English Wikipedia articles, and it answers noticeably better3
1
.Quality is mixed. Asked about Mario, it gets the basics right: a small plumber who rescues a princess. Asked what Metroid is, it produced blank entries plus "Great fuel" and "Good luck"
2
3
. Asked to define a large language model, it told one X user it is a dark quiet spot in the ground that does not see a dark spot3
.
Source: XDA-Developers
Useful it is not, Edwards says: it knows about 8,000 words, is millions of times smaller than ChatGPT's own models, and gets things wrong, at times with full confidence
1
. The fact book can carry mistakes too1
. He does not mind the failures, telling The Register the nonsense is almost cute given the model's limited knowledge and playful interface1
.Related Stories
Anyone can run it in a browser or download it as a Super Nintendo game file
2
1
. Real hardware needs a flash cartridge that supports the Super FX chip, such as the FXPak Pro1
.One outlet frames this as part of a recent trend of squeezing language models onto inadequate hardware. Per that report, days earlier a separate project got a language model running off a floppy disk
2
. For Edwards it is also a continuation. He has spent the past year theorizing about what he calls tiny talking machines. Before this came Séance, a chatbot on the original Nintendo console with an even smaller model, which simulated talking to a ghost1
.Edwards says his coding agents advise him the hardware limit has been reached, but he disagrees and is working on new ideas to make tiny models more coherent. "It's up to human creativity to transcend those assumptions," he told The Register. "I have already come up with new ideas that I believe can make a tiny model like ChatGPT for SNES even more coherent, and I'm working on them now"
1
.Summarized by
Navi
[1]
[2]