# Tiny 2.1 Transformer Chatbot Runs on 486 DOS PC From Single Floppy Disk

- URL: https://www.archynetys.com/trend/2026-11-22/a-380kb-chatbot-fits-on-one-floppy-disk-and-boots-on-a-486
- Day: 2026-11-22 | Category: Technology
- Sources: 5 | Articles: 5 | Verify: verified (100%)

> Developer WollyMonjac released Tiny 2.1, a 367,000-parameter transformer model quantized to 380 KB that boots on a 486DX or Pentium running MS-DOS from a single floppy disk.

## What happened Developer WollyMonjac published Tiny 2.1, a miniature transformer-based question-answering model that runs natively under MS-DOS on 486DX or Pentium hardware. The entire program, including its 8-bit quantized weights and a text-mode interface with conversation history, neural monitor panel, F-key shortcuts, and an optional PC-speaker voice, fits on one 1.44 MB floppy disk. The model was written in Turbo C 2.0 and can also be run in DOSBox or other emulators by extracting the GitHub release zip and executing the binary. ## Key context and numbers - **Model size:** 367,000 parameters across four layers, each 96 units wide, processing text character-by-character. - **Disk footprint:** Approximately 380 KB for 8-bit quantized weights; total runtime memory requirement about 540 KB of conventional DOS memory. - **Training data:** Roughly 3,500 facts, each phrased in multiple variations. - **Toolchain:** Written in Turbo C 2.0; model generated by "Fable 5.2" and trained by "Opus 5.5" according to the project notes. - **License:** MIT-licensed with a training folder included so users can retrain on custom topics. ## Why it matters Tiny 2.1 demonstrates that the transformer architecture can scale down to the severe constraints of early 1990s PCs — 640 KB conventional memory, no GPU, single-core integer-only CPUs — while retaining a functional chat interface. It serves as a pedagogical artifact showing the algorithmic core of modern LLMs stripped to its minimum viable form, and it highlights how quantization and tiny parameter counts can make AI run on virtually any von Neumann machine. The project also underscores a recurring limitation: outside its narrow training set the model hallucinates confidently, a behavior that persists regardless of model scale. ## What to watch next Community forks may port the inference engine to other retro platforms (e.g., 68k Macs, Amiga, early Windows 3.1) and experiment with larger context windows or alternative quantization schemes within the 640 KB barrier. The included training pipeline could spur hobbyist datasets for niche domains (vintage hardware repair, retro coding references). Meanwhile, the "machines making machines" provenance — a large LLM writing the C code for a tiny transformer — may inspire further research into automated model compression for edge and embedded targets.

## Sources Consulted
- [hackster.io — Bringing AI to the DOS Era](https://hackster.io/news/bringing-ai-to-the-dos-era-7a21c15723e6)
- [trusence.com — Developer fits a small question-answering transformer onto one floppy for DOS PCs](https://trusence.com/stories/developer-fits-a-small-question-answering-9600022c)
- [hackaday.com — Not-So-Large Language Model Fits On A Floppy Disk](https://hackaday.com/2026/11/18/not-so-large-language-model-fits-on-a-floppy-disk)
- [notebookcheck.net — Tiny AI runs on 486 DOS PCs from a floppy disk](https://notebookcheck.net/Tiny-AI-runs-on-486-DOS-PCs-from-a-floppy-disk.1456755.0.html)
