NES CPU Compiled into Qwen3.5 Model Weights, Running Games via Token Prediction

NES CPU Integration into Qwen3.5

The core of a Nintendo Entertainment System (NES), specifically its 6502 CPU and 2KB of work RAM, has been compiled into the weights of a stock Qwen3_5ForCausalLM model. This means the computational logic of the NES CPU is embedded within the neural network's parameters, rather than being emulated by traditional software.

This compilation was achieved without training the weights; instead, a compiler computed each weight directly. The model then operates by predicting the CPU's execution log one token at a time as a game runs. A separate host program manages the console's other functions, such as drawing the picture, reading controller inputs, and firing vblank interrupts.

Building on Previous Work

The nesformer project extends the work of Qwenputer by Divyansh Agrawal, which previously compiled CHIP-8 and RISC-V CPUs into Qwen3.5-4B models. Nesformer reuses Qwenputer's compiler and its method of using the model's Gated DeltaNet state as the machine's memory.

The project confirms that every CPU token generated during these runs originated from the model. For the included test games, every frame's output precisely matches that of a conventional 6502 emulator.

Technical Implementation and Usage

Users can set up nesformer by unzipping the project files and weights into the same folder. The CLI tool then allows running NES ROMs, with options to display frames in the terminal and save animated GIFs of the gameplay. The system requires Python 3.11 or newer and specific libraries like numpy, scipy, numba, and pillow.

Execution is noted to be slow, with initial setup taking about 20 seconds for engine compilation. Subsequent frames take 1–6 seconds each, meaning 10 seconds of NES gameplay can take several minutes to generate.