Show HN: A working Nintendo in ~1GB of Qwen weights (github.com)

A groundbreaking project named "nesformer" has successfully compiled a fully operational Nintendo Entertainment System (NES) within the weights of a Qwen3.5 AI model, requiring just 1GB of storage. This innovative approach leverages a compiler to generate the NES's 6502 CPU architecture, including its RAM and program ROM, by predicting the CPU's execution log on a token-by-token basis. The nesformer builds on a previous project called Qwenputer, which already demonstrated the capability to emulate simpler systems like CHIP-8 and RISC-V. This advancement allows enthusiasts to run NES games such as Super Mario Bros. using the model's weights, reproducing the CPU's operations with high accuracy while drawing graphics and reading inputs from a host program. This development is significant for the AI/ML community as it illustrates the potential of large language models to execute complex tasks beyond traditional text generation—essentially bridging AI with retro computing and game emulation. The nesformer achieves this by utilizing a unique memory management technique within the Gated DeltaNet architecture, effectively treating the model's state as the machine's RAM. This approach not only enhances the capabilities of language models but also showcases the versatility of transformer architectures in the realm of hardware simulation, opening up new avenues for research and experimentation in machine learning applications and retro gaming technologies.