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

The checkpoint has 466.6M parameters, only 9.3M of them nonzero, and zips to 3.3 MB. A compiler, not training, encodes a full NES 6502 CPU and 2 KB of RAM directly into the weights of a stock Qwen3.5 checkpoint. The model predicts the CPU's execution log token by token while a small host handles graphics, controller input and vblank timing. Throughput is about 1,650 tokens per second, against the roughly 700,000 a real NES needs, which makes it 150 to 290 times slower than the console. The method extends Qwenputer, a project by Divyansh Agrawal that compiled CHIP-8 and RISC-V CPUs into Qwen3.5-4B. Each Gated DeltaNet head keeps a 128x128 state matrix updated by the delta rule at every token. Coverage is narrow: it only supports NROM (mapper 0) cartridges and official 6502 opcodes. Interrupt handling is also limited, with no IRQs beyond BRK and NMI. For teams building AI products, this is a research demonstration, not a deployable technique. It shows that transformer weights can be engineered to run deterministic programs, which bears on interpretability and on hybrid systems that pair a language model with exact code. The speed gap and narrow compatibility mean it does not change near-term infrastructure choices, but it is a useful probe of how much exact computation weights can hold.