Oriel — AI & technology worth noticing, every day.

A Working Nintendo in ~1GB of Qwen Weights

Repository · https://github.com/dbreunig/nesformer

First seen by Oriel on .

AI summary

nesformer compiles an NES 6502 CPU, RAM, and NROM program ROM into a stock Qwen3.5 checkpoint, whose generated tokens reproduce CPU execution without trained weights.

  • A host program handles rendering, controller input, and interrupts; the model stores machine state in Gated DeltaNet memory and emits an execution log one token at a time.
  • The 466.6M-parameter bf16 checkpoint has 9.3M nonzero parameters and compresses to a 3.3 MB weights zip; it loads through stock Hugging Face transformers without custom code.
  • Tests report identical traces and frames for bundled games, including 600 Thwaite frames and 420 Full Game frames; the reference CPU also matches nestest and 5.7 million game-code instructions.
  • The default numba engine runs about 1,650 tokens per second on one CPU core, but the project is still 150–290× slower than a real NES.
  • Support is limited to NROM mapper 0, official opcodes, ROM-resident code, and approximate scanline rendering; there is no sound or IRQ support.

#qwen3.5 #nes #6502 #deltanet

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Points

Score reported by the source.

  • Hacker News: 283, latest observation ; +268 over observed period since

Comments

Comment count reported by the source.

  • Hacker News: 144, latest observation ; +125 over observed period since

Aliases

SourceExternal IDURL
Hacker News50306442https://github.com/dbreunig/nesformer Read discussion