ChatGPT fits in a Super Nintendo cartridge
A prototype runs a miniature version of ChatGPT on a Super Nintendo cartridge. The project, by Benj Edwards, is more of a retro demonstration than a useful assistant.
Benj Edwards, a tech journalist and retro enthusiast, has released ChatGPT for SNES, a miniature version of ChatGPT capable of running on a Super Nintendo cartridge, either emulated or on adapted hardware. The project mainly serves as a technical demonstration: the model is tiny, slow, and far from the reliability of modern versions, but it shows how far compressed AI models can go. At this stage, the feat is more of a wink than a daily tool.
Une IA de 2,2 millions de paramètres sur une console 16 bits
The core of the project is based on a model derived from GPT-OSS-120b, reduced to 2.2 million parameters according to the project site. To put the scale in perspective, Edwards explained to The Register that it is a “distilled” version, meaning condensed from a larger model to retain some of its capabilities with much less computation. The result: the system can generate text, but it often makes mistakes and can respond confidently, which remains a challenging skill in AI as in Monday morning meetings.
Two versions are offered. The first works only with the model, without a documentary base. The second adds a “fact book” built from the first sentences of about 150,000 simple English Wikipedia articles, which improves some answers but does not eliminate errors. Edwards also notes that the base version weighs about 4 MB, while the enriched version goes up to 19 MB.
Le projet joue avec les limites du matériel, pas avec la précision
To run ChatGPT for SNES, you need a device capable of running custom ROMs and managing the Super FX coprocessor. Edwards specifically mentions the FXPak Pro and SD2SNES cartridges. The version with the factual base also requires support for the MSU-1 system, an unofficial extension that adds CD-ROM-like functions to the Super Nintendo.
A “pure” version can even run without the Super FX chip, but Edwards notes that the responses are then less good. Its purpose is therefore not to compete with a modern assistant, but to test what a 1990s machine can still handle when asked to run a contemporary AI. The project is part of a series of experiments the researcher calls “tiny talking machines,” including an NES ROM called Séance, which simulates a conversation with a ghost in a haunted cartridge.
Ce que Benj Edwards cherche vraiment à montrer
Edwards says he designed the software scope, the interface, the chips used, and even the typing sounds, while the code was produced with the help of AI agents. According to him, this type of human-machine collaboration works best when the human retains creative direction. He considers having reached a practical limit with what such a small model can produce on such constrained hardware, but he continues to look for ways to make the whole more coherent.
The project will change neither the AI market nor the retro console market. However, it reminds us of a useful fact: model size remains a decisive parameter, and compression does not work miracles. Sometimes it creates wonders, sometimes very confident but off-target answers.
- 2.2 million parameters for the reduced model.
- 150,000 simple English Wikipedia articles used in one version.
- 4 MB for the base version, according to the project.
- 19 MB for the version with fact book, according to the project.
- Benj Edwards is the author of the prototype and its release.