ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsQuick take2 min readDisconfirmAI Editorial

Tiny 2.1: A 367K-Parameter Transformer Runs on a Floppy Disk for 486DX and Pentium PCs

A developer has created Tiny 2.1, an ultra-small question-answering transformer that fits on a floppy disk and runs on DOS-era hardware with just 540KB of memory.

2026-11-22

Developer WollyMonjac has released Tiny 2.1, a tiny question-answering transformer written in Turbo C 2.0, designed to run on 486DX and Pentium PCs under DOS. The model is not a large language model (LLM) but shares the transformer architecture, featuring only 367,000 parameters across 4 layers with a width of 96. It reads and writes letter by letter.

The entire project fits on a standard 1.44MB floppy disk, utilizing the limited space efficiently. The model weights are quantized to 8-bit integers (one scale per row), totaling approximately 380 KB. The system was trained on about 3,500 facts, each presented in various wordings. To run, it requires about 540KB of free conventional memory on the host machine.

Tiny 2.1 includes a clean user interface with a conversation history pane, a neural monitor panel, and an 'Ask Tiny' prompt window, along with F-key shortcuts typical of DOS programs. A notable feature is its ability to speak responses using the PC speaker, producing a retro-futuristic robot voice.

The GitHub repository states that the code was "Written by Fable 5.2 and trained by Opus 5.5." The developer acknowledges that outside its specific training data, the model "makes things up confidently," a behavior reminiscent of LLM hallucinations but framed here as a fun feature. The project is MIT-licensed and includes a training folder, allowing users to potentially build custom question-answer transformers tailored to specific topics.

Sources

Sign in / Sign up
ReadingResearchRadarInvestment framework
tomshardware.comTom's Hardware↗
Continue researching →