Software engineer Yeo Kheng Meng has created a version of a large language model (LLM) that is capable of running on DOS-based computers, including machines with 486 and Pentium 1 processors.
To achieve this, Meng implemented Andrej Karpathy’s Llama2.c library, which he had previously used to run the LLM on Windows 98.
While the processors are no longer powerful enough to run most modern applications, Meng found that they were capable of running certain smaller LLMs.
For instance, a 486 computer was able to run a 260 kB model that output two tokens per second.
However, bigger models are also available, with one 110 Mb model running faster on a Pentium M Thinkpad T24 than on Meng’s modern Ryzen 5 desktop, owing to a memory allocation error on the latter.
The code has been adapted to run on any 32-bit i386 hardware, with a 16-bit version the next challenge.
