nesformer

nesformer is an open-source project by Drew Breunig (GitHub dbreunig), created on October 30, 2026, that compiles a Nintendo Entertainment System CPU into the weights of a stock Qwen3.5 language model. The NES's 6502 CPU (core of the 2A03 chip), its 2 KB of work RAM and a cartridge's program ROM all run inside a standard Qwen3_5ForCausalLM checkpoint.

No training is involved — a compiler computes every weight. When you play, the model predicts the CPU's execution log token by token, while a small host program draws the picture, reads the controller and fires the vblank interrupt. Demos include Super Mario Bros. World 1-1.

How it works

• Built on Qwenputer: it reuses the compiler from Divyansh Agrawal's Qwenputer project (which compiled CHIP-8 and RISC-V CPUs into Qwen3.5-4B).
• Central trick: the model's Gated DeltaNet recurrent state is the machine's memory, so the transformer literally holds RAM in its hidden state.
• Exactness: for the bundled test games, every frame matches a conventional 6502 emulator's output exactly.
• Division of labour: the LLM executes the CPU; a lightweight host handles graphics (PPU), input and interrupts.

Why people care

It is a vivid demonstration that modern LLM architectures are general computers: you can place a real, deterministic program inside the weights without gradient descent. That makes it interesting for interpretability, "compiled" neural programs and research on giving LLMs exact algorithmic sub-modules — and it is simply a fun hack (Mario running on an LLM).

Try it

1. Get the code and weights from github.com/dbreunig/nesformer (unzip nesformer.zip and nesformer-weights.zip into the same folder, with the weights next to the nesformer/ package).
2. Follow the README quick start to run a bundled test ROM; bring your own legally obtained ROM for other games.
3. Compare output with a normal 6502 emulator to see the exact match.
4. For background, read Divyansh Agrawal's Qwenputer write-up (divagr.com/blog/qwenputer).