Squeezing a language model onto a 1975 8-bit chip

Someone got a modern AI language model running on an 8-bit processor from 1975 with only 32KB of RAM.

Developer Matt Beton wrote a custom inference engine to run a tiny Mamba-based language model on the MOS 6502 CPU—the chip inside his dad's 1980s BBC Micro. The entire setup fits in 22KB of memory: 9KB for code and 13KB for 52,000 model parameters. Because the 6502 lacks a multiplication instruction, Beton used BitNet quantization—a technique that shrinks weights to -1, 0, or 1—to swap costly multiplications for quick additions, then transferred the code using a DIY headphone-to-tape cable.

Why it matters: It shows how far extreme compression can stretch ancient hardware. By pairing ternary weights with a small 27-character vocabulary, modern neural network architectures can execute logic on silicon built decades before transformer models existed.

Try it: You can run the model in your browser using an emulated BBC Micro that auto-loads the tape image from GitHub. Text generation takes a few minutes.

The generated story about Tom and Lily won't win a Pulitzer, but teaching 50-year-old hardware to predict text is an impressive feat.