We’ve spent the last few years treating Large Language Models (LLMs) as chatbots or creative writers. But what happens when you stop treating a model as a conversationalist and start treating it as a piece of hardware? Enter Nesformer, a wild experiment by developer dbreunig that manages to cram a fully functional Nintendo Entertainment System into about 1GB of Qwen3.5 weights.
Not Training, But Compiling
Here is the mind-bending part: this isn't a model that was "trained" to play Mario by watching thousands of hours of gameplay. Instead, it uses a compiler to encode the actual logic of the NES 6502 CPU, its 2KB of RAM, and the game's ROM directly into the Gated DeltaNet state matrices of a stock Qwen3.5 checkpoint.
Essentially, the model isn't guessing what happens next in the game; it is predicting the CPU's execution tokens one by one, effectively acting as the hardware itself.
From Tokens to Tetris
By utilizing roughly 466.6M bf16 parameters (with only 9.3M being nonzero), Nesformer proves that the weights of an LLM can serve as a functional world model or emulator. When you press a button, the model processes that input and predicts the resulting state of the CPU. The result? A working version of Super Mario Bros. running not on a chip, but inside the mathematical weights of a transformer.
The Future of 'Weight-Based' Computing
While playing Mario in an LLM might seem like a high-effort party trick, it hints at a deeper shift. If we can compile legacy hardware into weights, we are moving toward a future where software isn't just run by AI, but is encoded within it. We're seeing the first steps toward a hybrid of traditional deterministic computing and neural networks.
Sources
Media
