A developer has released Tiny 2.1, a language model designed to run on 486DX and Pentium-era hardware running DOS. The project, created by WollyMonjac, fits onto a standard floppy disk, which typically holds just under 1,475,000 bytes of data. Unlike modern large language models, Tiny 2.1 is described as “a tiny question-answering transformer written in Turbo C 2.0.”
The model utilizes a transformer architecture with 367,000 parameters organized into four layers that are 96 wide. To accommodate the limited storage of a floppy disk, the software uses 8-bit weights with one scale per row, totaling 380 KB. The system was trained on approximately 3,500 facts, with each fact presented in multiple wordings to facilitate question-answering.
Running the software requires approximately 540KB of free conventional memory on legacy hardware. The user interface includes a conversation history pane, a neural monitor panel, and an Ask Tiny prompt window. The program also features a PC-speaker robot voice capable of speaking the model's responses.
According to the GitHub repository, the project was “Written by Fable 5.2 and trained by Opus 5.5.” Monjac noted that the model is prone to errors, wryly admitting that outside of its specific training, “it makes things up confidently.”
The distribution is released under an MIT license and includes a training folder for users who wish to build a custom question-answer transformer. This allows individuals to tailor the model to specific topics by utilizing the provided training data.














