Show HN: A working Nintendo in ~1GB of Qwen weights

Category: ai-ml

Tags: llm, emulation, compiler, nintendo-nes, qwen

Score: 7.8/10 (Innovation: 9, Technical: 10, Documentation: 8, Utility: 4)

nesformer compiles an NES CPU (6502) into the weights of a stock Qwen3.5 model, using the model's Gated DeltaNet state as machine memory. It runs NROM games by predicting the CPU's execution log token-by-token, with a host program handling graphics, input, and interrupts. The project is a novel demonstration of in-weights computation and achieves exact frame matching with a reference emulator.

Target audience: AI researchers, emulation enthusiasts, machine learning engineers

Repository: https://github.com/dbreunig/nesformer · Python · NOASSERTION · 239 stars

View on Hacker News