Trusence Every claim has a source
Last updated 22 November 2026 Search Türkçe
Research

Developer fits a small question-answering transformer onto one floppy for DOS PCs

The model runs on 486DX and Pentium machines and can be retrained by users on their own topics, though its author warns it invents answers outside its training.

A developer known as WollyMonjac has built Tiny 2.1, a small question-answering transformer written in Turbo C 2.0 that fits on a single floppy disk and runs on 486DX and Pentium PCs under DOS. It is not a large language model but shares the transformer architecture, with 367,000 parameters, 8-bit weights totalling about 380 KB, and training on roughly 3,500 facts each phrased in many ways. The program needs around 540 KB of free conventional memory and offers a conversation history pane, a neural monitor panel, F-key shortcuts and an optional PC-speaker robot voice. The author openly warns that the model makes things up confidently outside its training, and the MIT-licensed distribution includes a training folder so users can tailor it to their own topics.

Why it matters

The project shows that a transformer model can be small enough to run on very old hardware under DOS, with a memory footprint of about 540KB of free conventional memory. For people who follow computing history, it is a concrete example of how little storage and memory the architecture needs when the model is trained on a narrow body of material. Its limits matter as much as its size: the author notes it answers confidently even when a question falls outside what it was trained on, so it is a demonstration rather than a dependable source of information.

Sources · 1