The intersection of modern artificial intelligence and obsolete vintage hardware continues to yield astonishing engineering feats. While developers and enthusiasts have increasingly experimented with pushing lightweight language models onto restrictive, unconventional hardware—ranging from retro game consoles to a modified 2015 Kindle—a new achievement has taken retro-computing compatibility to an entirely different era. A developer has successfully engineered a functional language model that fits on a standard floppy disk and can be booted directly onto a DOS-compatible vintage computer.

The project, which has quickly gained traction within online tech communities, demonstrates how cutting-edge AI tools can be leveraged to build software for architecture that predates large language models by decades.

Getting a Language Model to Run Off an Intel Pentium is No Easy Feat

Over on the ClaudeAI subreddit, a user operating under the handle wollymonjak shared the details of this unusual project. Rather than manually writing and training a model from scratch using modern data pipelines alone, the creator utilized advanced AI assistance to streamline the process. Specifically, the developer used Claude Fable 5.2 to help design and build the underlying transformer architecture, while employing Opus 5.5 to assist in supplying the necessary training data.

The result is a remarkably compact system that defies conventional expectations of modern AI infrastructure. The entire language model fits neatly onto a traditional floppy disk, allowing it to boot smoothly on vintage DOS hardware. Running an artificial intelligence model on a CPU architecture like an Intel Pentium—designed long before transformer networks, neural processing units, or large-scale machine learning frameworks were even conceptualized—is a formidable technical hurdle. The sheer difference in computing paradigms makes running modern-adjacent AI on such legacy systems an exercise in extreme optimization and clever software engineering.

Someone used Claude to build an AI chatbot for MS-DOS, and it fits on a floppy

But Someone Got It Working Anyway

Despite the immense hardware limitations, wollymonjak managed to get the system operational at a respectable speed, given the decades-old silicon driving it. Because of the extreme constraints imposed by the floppy disk format and the vintage DOS environment, the resulting AI is naturally quite limited in its capabilities.

The developer has candidly noted that the model is not designed for heavy lifting. When prompted with inputs that are overly complex or too long for the tiny network to process, the system quickly degrades and begins generating erratic nonsense rather than coherent text. However, functional limitations do little to diminish the novelty and technical impressiveness of the achievement. Successfully booting and interacting with any form of a language model within a standard DOS environment is an accomplishment worthy of high praise among retro-computing and AI hobbyists alike.

For those interested in exploring the inner workings of the project, the creator has made the code, specifications, and project files publicly available. Enthusiasts and curious developers can examine how the transformer was constructed, review the training methodology, and even download the disk image to test it out for themselves via the project’s official GitHub page under the moniker Tiny-2.1.