It is a working reminder that a binarised network fits in 20 KB and needs no multiply instruction, which matters if you ship models to tiny hardware.

MNIST64 is a binarised convolutional network for digit recognition that runs on a stock 1982 Commodore 64: one megahertz, 64 KB of RAM, no multiply instruction. Jarno Heikkinen wrote the inference in hand-tuned 6502 assembly, fitting the model in 20,852 bytes and recognising a digit in about 0.65 seconds. A confidence gate lets easy digits exit early. On 50,000 MNIST and EMNIST images it reaches 99.42%, edging past the 1998 LeNet result. The repo holds training code, the 6502 implementation, a disk image and a paper, and you can draw digits with a mouse or joystick. No licence is stated.

More like this

Doomformer

Compiles small neural networks into DOOM maps you can walk through

jdBasic

A modern BASIC where you pause a running program, change the code and carry on

bootgpt

512 bytes of assembly that boot a PC straight into a neural net

Jinfer

An AI inference engine that runs inside the JVM, with no Python sidecar

PicoLM

A C99 LLM engine that streams weights from disk, from Tru64 to Raspberry Pi

MAPLE

Gemma 4 26B cut to 48 experts per layer so it fits a 6 GB laptop GPU