
35-Year-Old SNES Runs ChatGPT With Super FX 2 Acceleration
Benj Edwards has built an unofficial ChatGPT app for the Super Nintendo, using a 2.2-million-parameter model and the Super FX 2 coprocessor to speed replies on vintage hardware.
Nintendo's classic Super Nintendo Entertainment System now has an unofficial ChatGPT chatbot, created by Vintage Computing blogger Benj Edwards. The project uses a 2.2-million-parameter distillation of OpenAI's GPT-OSS-120b and accelerates responsiveness with the Super FX 2 coprocessor. Edwards describes the result as fun, though in Tom's Hardware's own testing its usefulness beyond curiosity was very limited.
What the chatbot offers
The app can be tried through an embedded SNES9x emulator. Users can choose a 4MB model or a 19MB model that includes the opening sentences of 150,000 Simple English Wikipedia articles in its knowledge base, stored in the cartridge's MSU-1 data file. The larger option is recommended so the model knows more facts and can work as a calculator.
After loading, the app asks the user to press Start; on a PC keyboard, that means pressing Enter. Text entry does not require a joypad, and typing works while the window is in focus. The app also includes SNES-appropriate input sound effects and a chirpy chip-tune soundtrack. A tweet dated November 19, 2026, shows the SNES app splash screen, chat window, and an animation of someone interacting with the chat.
Why Super FX 2 matters
Edwards used the Super FX 2 chip to accelerate the language model. The coprocessor enhanced SNES games with faster 2D and 3D graphics and debuted with Yoshi's Island. It contains a 21.47 MHz RISC CPU, compared with the SNES main CPU, a 16-bit Ricoh 5A22 running at 2.68 MHz. Nintendo's second version of the chip retained the original's polygon and advanced 2D effect capabilities, ran at full speed internally rather than half speed, and supported larger ROM sizes.
Running it on original hardware
The project is not only an online diversion. It can be downloaded and run on a real vintage SNES, but that requires a flash cart with Super FX support. MSU-1 support is also needed for the build that includes the abridged Wikipedia fact book.
Edwards acknowledges the model's limits, saying it is millions of times smaller than the models behind ChatGPT. With a small and fuzzy knowledge base, he says, "You'll get some fun and weird stuff."
Sources: Tom's Hardware
SiTech — AI-powered web development
We build fast, modern websites and bring AI into real business workflows. Have a project or a question? We'd love to help.