A ChatGPT-style chatbot now runs on a real Super Nintendo. Benj Edwards, a tech writer, built it with help from GPT-6 Astra and the open-weight model gpt-oss, producing a 2.2-million-parameter model that answers questions directly on the 1990s console. Edwards calls the project unofficial and not endorsed by OpenAI. He posted it on X on November 19, where it has drawn over 317,000 views.
The console's graphics chip
The starting idea was simple: AI models run on GPUs, so Edwards wondered how well a chatbot would run on the most powerful chip in the console. He chose the Super FX 2, the coprocessor Nintendo placed inside some cartridges to boost the SNES's graphics and computing power. The model is tiny, orders of magnitude smaller than the cloud services that run models with hundreds of billions of parameters.

Image: benjedwards — https://x.com/benjedwards/status/2123457434962608525
According to the project documentation on BX Foundry, the current version is 0.7, dated November 2, 2026. The training conversations were written by gpt-oss-120b, OpenAI's open-weight model, while the program's code was produced by AI agents. Published images show the title screen on a CRT monitor and a chat interface answering questions such as "who is Mario."
Stated limits
Edwards admits the system makes mistakes. He describes it as a playful experiment rather than a usable tool. Its value lies in the contrast with today's giant models: how much computing power a conversational assistant really needs, and where the limit of the hardware begins. Similar comparisons have emerged in other console experiments, such as a jailbroken PS4 running a 9-billion-parameter model offline and a PS5 running a 27-billion-parameter model with heavy compromises. Old consoles, in other words, are becoming testing grounds for AI.
The project also used GPT-6 Astra, OpenAI's flagship launched in September, already presented as the start of the AGI era. Edwards used it as a development tool, not as a component of the chatbot running on the console.
The case shows how far AI software can shrink: a model with thousands of times fewer parameters than those in data centers can still hold a conversation, albeit with errors. The SNES, released in Japan in 1990, was never designed for anything like this, which is why the result draws attention from retro computing and AI enthusiasts alike.

No comments yet. Be the first!