Can You Play Super Mario Inside LLM Weights? Meet Nesformer

According to Simon Willison's weblog covering Drew Breunig's repository, developer Drew Breunig successfully embedded a working Nintendo Entertainment System 6502 CPU emulator directly into the weights of the Qwen3.5-4B model. Inspired by previous work like the 'qwenputer' project, this experiment, named 'nesformer,' allows the model to execute actual NES ROMs.

While playing Super Mario Bros directly through neural network weights sounds like science fiction, the reality comes with severe performance caveats. Because the emulation logic lives inside the transformer weights rather than native compiled machine code, the gameplay runs at a fraction of standard speed, making it more of an impressive technical novelty than a practical gaming setup.

How it works in practice

To build the system, Breunig used Claude (Opus 5.5) prompts to explore if a chip and RAM emulation could be baked into model weights using techniques similar to hardware FPGAs. Opus even tracked down a 15-year-old NES CPU conformance test suite by Christopher Pow to validate the logic.

• Clone or download the nesformer repository provided by Drew Breunig.
• Configure the command-line interface (CLI) to point directly at your desired NES ROM file.
• Run the Qwen3.5-4B modified weights locally to execute the instructions.
• Expect extremely sluggish execution speeds as the neural network processes the CPU emulation cycles.

Is it worth your time?

As a tool for your daily software engineering workflow, nesformer holds zero practical utility. It will not accelerate your coding or optimize your deployment pipelines. However, for developers curious about the boundaries of model weight manipulation and prompt-driven engineering, it offers a fascinating glimpse into unconventional computing paradigms.

Inspired by qwenputer by Divyansh Agrawal, Drew Breunig baked a working emulator for a Nintendo Entertainment System 6502 CPU into the weights of Qwen3.5-4B.

FAQ

• Can I play games at normal speed using nesformer? No, running an emulator inside LLM weights is currently extremely slow and intended purely as an experimental proof of concept.
• What model does nesformer use? The project modifies the model weights of Qwen3.5-4B to handle the 6502 CPU instruction set logic.
• Do I need a special GPU to run it? You need a local environment capable of running the Qwen3.5-4B model along with the repository's custom CLI tool.