Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL

WollyMonjac/Tiny-2.1

Machine Learning

A 367K-parameter question-answering transformer in Turbo C for DOS, small enough to fit on a floppy disk.

106
14
C
new today

What it is

TINY 2.1 is a small question-answering transformer implemented in Turbo C 2.0 and built to run under plain DOS on a 486DX or Pentium. The model has 367,000 parameters across 4 layers with width 96, uses 8-bit weights with one scale per row, and occupies about 380 KB, so the program and model fit on a floppy disk. It answers by reading and writing one character at a time and includes a full-screen interface with a live neural monitor and PC-speaker voice output.

The model was trained on roughly 3,500 facts, each phrased in many ways. The README explicitly warns that it confidently fabricates answers outside that training set, which matters when deciding whether it fits a use case.

Run it

  1. Prepare a DOS machine or DOS environment with about 540 KB of free conventional memory.
  2. Place TINY.EXE and TINY.BIN in the same folder, such as C:\TINY.
  3. Start the program:
C:\> CD \TINY
C:\TINY> TINY

To rebuild from source, run TCC -ml TINY.C in the same directory using Turbo C 2.0.

Constraints

  • Requires a 486DX/Pentium-class DOS system and roughly 540 KB of conventional memory.
  • The executable needs both TINY.EXE and TINY.BIN side by side.
  • Answers are limited to the trained fact set; outside it, outputs can be confidently wrong.
  • The repository does not declare a license.

HotRepo's take

The repository first appears on HotRepo's new-project list on 2026-11-16 with 104 stars and 13 forks, all gained within one recorded growth window of less than a day. That is a single snapshot, so it supports interest at launch but says nothing about lasting adoption. The distinctive technical fact is the extreme resource envelope: a character-level transformer squeezed into about 380 KB of 8-bit weights and runnable on DOS-era hardware, with README-provided build instructions, rather than a modern general-purpose model.

Project source and documentation

Written by AI from the project's README, updated . The project's own docs are the source of truth.

You might also like

  1. n4ze3m/tiny-vocaA 4.5M-parameter English text-to-speech model with two voices that runs on CPU or in the browser.Python105
  2. JustVugg/colibriRun very large MoE models on consumer hardware by streaming experts from disk, with CPU, Vulkan, CUDA and Metal backends.C46,109