nesformer Packs Nintendo Into Qwen Weights

This is a wonderfully impractical but technically serious demonstration that transformer weights can serve as executable digital hardware, not merely approximate it. The compelling result is the compiler-and-state design; the sobering result is how badly general-purpose neural computation compares with a dedicated 8-bit CPU.

• –The roughly 933 MB bf16 checkpoint contains 466.6M parameters, but only 9.3M are nonzero because a compiler analytically emits the circuit rather than training it. [GitHub](https://github.com/dbreunig/nesformer)
• –Qwen’s Gated DeltaNet state becomes addressable memory for registers, RAM, ROM, and execution bookkeeping, extending the earlier Qwenputer approach. [Qwenputer](https://github.com/divagr18/qwenputer)
• –Validation is unusually strong for an LLM demo: the project reports exact traces and matching rendered frames across multiple test runs.
• –The practical scope is narrow: only NROM mapper games are supported, while PPU, controller, and most audio behavior remain outside the weights.
• –For AI developers, nesformer points toward compiled deterministic subsystems inside model architectures, though current speed makes it research art rather than a viable emulator.