Llama 2 LLM on DOS (2025) (yeokhengmeng.com)

A developer has successfully ported the Llama 2 Large Language Model (LLM) to run on vintage PCs operating under DOS, demonstrating the capabilities of local LLM inference on machines as old as the Toshiba Satellite 315CDT from 1996. By utilizing the llama2.c project, originally created by machine learning expert Andrej Karpathy, the developer recompiled the LLM for older systems without high-performance specifications, challenging the prevailing idea that LLMs necessitate powerful, modern hardware. The project is open-sourced, allowing retro computing enthusiasts to explore and experiment with this innovative integration. The significance of this achievement lies in its potential to democratize access to AI capabilities, enabling users with limited resources to run LLMs locally without reliance on cloud services. Technical intricacies included adapting floating-point operations and memory management for DOS compatibility, alongside using specific DOS extenders to leverage higher memory capacity. Benchmarks show impressive inference speeds, notably with the ThinkPad T42 outperforming newer models in certain aspects, underscoring the continued relevance of retro hardware in the modern AI landscape. This breakthrough inspires further exploration of AI applications on resource-constrained systems.