Sulat.com / Writing on AISunday, November 22, 2026
Artificial Intelligence/22 November 2026

ChatGPT for SNES runs a 2.2-million-parameter model on the Super Nintendo’s Super FX 2 chip

Benj Edwards's unofficial ChatGPT for SNES runs a small model trained on gpt-oss-120b conversations from a Super Nintendo cartridge. The project page says it isn't made or endorsed by OpenAI.

Green 3D OpenAI logo standing on a grey game cartridge plugged into a retro console, with a CRT television showing a chat window

At a glance

  • Benj Edwards's unofficial ChatGPT for SNES runs a language model of about 2.2 million parameters on the Super FX 2 chip inside a Super Nintendo cartridge, writing answers one word at a time.
  • The model learned from training conversations written by gpt-oss-120b, OpenAI's open-weight model, and Edwards's project page says the project is not made or endorsed by OpenAI.
  • Version 0.7 has a pure build that loads about 4 MB in the browser and a fact book build that adds 150,000 Simple English Wikipedia openers and loads about 19 MB.
  • On a real console, both builds need a flash cartridge with Super FX support such as an FXPak Pro or sd2snes, and the fact book build also needs MSU-1 support.
  • Edwards designed the project and AI coding agents wrote the code; he told The Register he has new ideas for making the tiny model more coherent.

Benj Edwards has built an unofficial ChatGPT for SNES, a chat program whose language model of about 2.2 million parameters runs on the Super FX 2 chip inside a Super Nintendo cartridge.

It isn’t the ChatGPT service running on a console. The model is a small one that learned from conversations written by gpt-oss-120b, OpenAI’s open-weight model, and Edwards’s page says the project is “Not made or endorsed by OpenAI.” He told The Register it is a distilled version of gpt-oss-120b.

Edwards announced it on X on 19 November, crediting GPT-6 Astra and gpt-oss for help with the build.

Benj Edwards on X, 19 November 2026: the announcement of ChatGPT for SNES.

How the Super FX 2 runs a language model

The Super FX 2 is a 21 MHz coprocessor that debuted in 1995 with Yoshi’s Island, where it scaled and rotated sprites. An earlier version of the chip drove the 3D graphics in Star Fox in 1993, Edwards notes.

Benj Edwards on X, 19 November 2026: the Super FX 2 chip behind the model.

Here the chip runs the model, which writes its answers one word at a time. How-To Geek reports that, per Edwards, the chip reads the model’s 8-bit weights straight from the cartridge ROM and generates about two tokens a second, or roughly one word every half second.

The chip isn’t strictly required. The pure build could run from a real custom cartridge without the chip, Edwards said, and he has a version with no Super FX chip at all, though the answers aren’t as good without the extra processing power.

Four stat cards: 2.2M parameters, 21 MHz Super FX 2, about 8,000 words known, about 2 tokens a second
A 2.2-million-parameter model, a 21 MHz chip and roughly two tokens a second.

Two builds, in a browser or on a real console

Version 0.7, dated 2 November, comes in two builds. In the pure build, every word comes from the model; the ROM is a 1.5 MB zip and the browser version loads about 4 MB.

The fact book build adds the opening sentences of 150,000 Simple English Wikipedia articles, stored in an MSU-1 data file. Like ChatGPT searching the web, the model can decide to look something up and print what the book says, and Edwards’s page says it also has a calculator. That ROM is a 17 MB zip and loads about 19 MB in the browser.

Side by side: pure build of 4 MB needing a Super FX flash cart, fact book build of 19 MB also needing MSU-1
The fact book build adds 150,000 Wikipedia openers and needs MSU-1 support on a real console.

On a real Super Nintendo, both builds need a flash cartridge with Super FX support, such as an FXPak Pro or an sd2snes, with the Super FX speed option left at Normal. The fact book build also needs MSU-1 support.

Trying it

Edwards’s page runs both builds in the Snes9x emulator inside a browser tab. You type on your keyboard, and each letter takes a moment to appear because the page moves the cursor across the game’s own on-screen keyboard. Asked who Mario is, the cartridge answers that he’s a small plumber who jumps and saves a Princess.

The page says the model knows about 8,000 words and is millions of times smaller than the models behind ChatGPT. When it slips up, Edwards said, “it’s almost cute when it does it because of its limited knowledge and its playful interface.”

Who made what

Edwards is a tech journalist who runs the Vintage Computing site. His page says AI coding agents wrote the code, and he described his own part to The Register.

I designed the scope of what the software should be, how it should work broadly, what chip it should use, how the interface should look and work, and the sounds it makes when it types responses.

Benj Edwards, to The Register
Four stacked boxes: Benj Edwards designs it, AI agents write the code, gpt-oss-120b writes training chats, a 2.2M model results
Edwards designed it, AI agents coded it, and an OpenAI open model wrote the chats the tiny model learned from.

He has a second project, Séance, an NES ROM that runs an even smaller model to simulate talking to a ghost in a haunted game cartridge.

The coding agents advise that the model has hit the limit of what the hardware allows, Edwards said. He doesn’t take their word for it, and has new ideas for making a tiny model more coherent that he’s working on now.

Sources

Disclosure: I’m an OpenAI, Firecrawl and Devin ambassador.


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.