2 Sources
[1]
Someone used Claude to build an AI chatbot for MS-DOS, and it fits on a floppy
* This Claude-built transformer runs on DOS from a floppy disk, booting on vintage Pentium hardware. * It has tiny specs, with 367K params, 4 layers, 8‑bit weights at ≈380KB, and needs ≈540KB of conventional RAM. * It has limited knowledge and hallucinates on complex queries, but a highly
[2]
Someone used Claude to build a chatbot that fits on a floppy disk
A lifelong musician, Hamed plays multiple instruments, including guitar, bass, and oud. He is equally fascinated by smart-home technology and can happily spend hours geeking out over connected lights, speakers, sensors, and automations that make everyday life a little more interesting. When he
Share
Copy Link
A chatbot that runs on 1990s-era computers and fits on a floppy disk was built with Claude. It works, within narrow limits.
A hobbyist developer known as WollyMonjac has built a chatbot that runs on Pentium and 486DX machines under DOS, Microsoft's old disk operating system, and fits on a single floppy disk
1
2
. It is a demonstration of AI models running on hardware they were never intended for, from a single floppy disk.
Source: MakeUseOf
The project was AI-built end to end. As both outlets report it, Claude Fable 5.2 wrote the transformer, the core structure of the model, and Claude Opus 5.5 generated the training data
1
2
. The result has 367,000 parameters. It sits orders of magnitude below mainstream language models . It runs in about 540 kilobytes of free conventional memory, roughly what DOS programs ever had to work with1
2
. It reads and writes text letter by letter, using about 380 kilobytes of 8-bit weights1
2
.What can half a megabyte of model actually do? It answers simple trivia, such as naming the capital of Peru, spells "cat" backward, does basic arithmetic and tolerates small typos
1
2
. On screen it is a neon full-screen display with a live neural monitor, plus a robot voice through the machine's internal speaker1
2
.
Source: XDA-Developers
The limits are hard. The model was trained on about 3,500 facts, each phrased in many ways, plus small talk and arithmetic examples
1
2
. Ask it anything outside that set and it confidently makes things up. In its own tests the creator reported about 90% accuracy on questions phrased like the training data, falling to 61% when the questions were reworded2
. Division trips it up; multiplication goes better.The project was showcased on the ClaudeAI subreddit, and the C source code and model are downloadable from GitHub
1
2
. One write-up places it in a wider hobbyist pattern of getting chatbots onto restrictive hardware, including a language model that ran off a 2015 Kindle1
. MakeUseOf's verdict is the honest one: it is a proof of concept and a weekend project, not something to rely on2
.Related Stories
Summarized by
Navi
14 Dec 2024•Technology

01 Nov 2024•Technology

12 Aug 2025•Technology
