News › 2025 › May 2025

Llama 2 inference on a Commodore 64

Project2025-05-18

Maciej Witkowiak ported Karpathy's llama2.c to a Commodore 64 with a 2 MB REU. It continues a tiny-stories prompt. It is not a chat model.

llama2.c64 is Maciej Witkowiak's port of Andrej Karpathy's llama2.c to the Commodore 64, compiled with oscar64. It needs a 2 MB RAM Expansion Unit and runs the 260K tinystories weights. Witkowiak is explicit: this is not a chat model. It continues a story the way a small language model does.

Under VICE, enable the REU at 2048 KB, point it at the supplied weights image, and load the PRG. Warp mode is how the author runs it.

Links