GoKawiilDeveloper compiles a working NES emulator into Qwen3.5 model weights
A project called nesformer encodes an NES's 6502 CPU, its 2KB of work RAM, and a game cartridge's program ROM directly into the weights of a Qwen3.5 language model, with no training involved—every weight is computed by a compiler. During gameplay the model predicts the CPU's execution log token by token, while a small host program handles rendering, controller input, and interrupts. The project builds on an earlier tool called Qwenputer, reusing its compiler and its technique of storing machine memory in the model's Gated DeltaNet state.
GoKawiil's interpretation of the reporting above, not reported fact.
This demonstrates that transformer architectures can be repurposed as general-purpose computation substrates rather than purely statistical predictors, since the model's internal state is being used to exactly emulate a real CPU's behavior cycle by cycle. The developer notes that output for test games matches a conventional 6502 emulator exactly, suggesting the approach is not just a novelty but a verifiably accurate alternative way to run legacy hardware logic inside modern AI model formats. It could hint at future directions where AI weights double as portable, verifiable emulation layers for other classic systems.
- nesformer compiles an NES's CPU, RAM, and ROM into the weights of a Qwen3.5 model with no training.
- The model predicts CPU execution logs token by token while a host program manages graphics, input, and interrupts.
- Output for bundled test games matches a standard 6502 emulator exactly, per the developer's verification process.
Source: github.com, 2026-10-30
Published there as: “A Working Nintendo in ~1GB of Qwen Weights”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.