Skip to main content
The Quantum Dispatch
Back to Home
Cover illustration for TINY 2.1: A DOS Transformer Chatbot That Fits on a Floppy

TINY 2.1: A DOS Transformer Chatbot That Fits on a Floppy

TINY 2.1 packs a 367,000-parameter transformer into 380KB of weights. It runs under DOS on a 486DX or Pentium, and you can retrain it with your own facts.

Alex Circuit
Alex Circuit★Nov 18, 2026★3 min read

Large language models keep getting larger, which makes TINY 2.1 a delightful bit of engineering mischief. This tiny transformer runs under plain DOS on a 486DX or Pentium PC, and the whole thing fits on a floppy disk. Created by developer WollyMonjac and spotted by Hackaday, the project is not going to replace a modern assistant. It is something more fun: a working demonstration of how far the core idea behind today's chatbots can be squeezed.

  • Model size: 367,000 parameters across four layers, 96 wide, reading and writing one letter at a time.
  • Storage: 8-bit weights totaling 380KB.
  • Memory: about 540KB of free conventional memory.
  • Knowledge: trained on about 3,500 facts, each asked in many different wordings.

How Does TINY 2.1 Run on a 486?

The numbers are what make this project sing. A 367K-parameter network is minuscule by current standards, and 8-bit weights with one scale per row keep the model file to 380KB. The program is written in Turbo C 2.0, the kind of toolchain that feels right at home beside a beige tower and a stack of 3.5-inch disks. Setup is equally old-school: copy the files from the repository's dos folder to the machine, change into the program directory and run TINY. The executable and TINY.BIN weight file need to sit in the same folder.

The target hardware is a 486DX or Pentium. Conventional memory is the constraint to plan around, since 540KB is a big request in the classic 640KB DOS world. Expect to do a little boot-configuration housekeeping if your system loads many drivers.

The presentation gets the retro treatment too, with a neon full-screen interface, a live neural monitor and a robot voice through the PC speaker.

What Can This Floppy Disk LLM Actually Do?

TINY 2.1 answers questions about the roughly 3,500 facts in its training set. Ask outside that range and it will confidently invent an answer, as the project's documentation plainly says. Hackaday makes a fair point about the naming here too: at this size, "transformer" is a more honest description than "large language model."

That limitation also reveals the project's most interesting feature. Because the knowledge base is small and explicit, you can decide exactly what the model should know. A retro computing club could build a museum guide, or a hobbyist could train one around a favorite game manual.

Can You Train TINY 2.1 Yourself?

Yes. The repository includes the facts, a generator that turns them into question-and-answer pairs and a PyTorch training script. You need Python 3, NumPy and PyTorch, and the author says a normal CPU is enough. Add facts to the training folder, train, then export a replacement TINY.BIN for the DOS executable to load.

There is a very 2026 twist in the credits: the README says the code was written by Claude Fable 5.2 and the model trained by Claude Opus 5.5. The code is MIT licensed, while three data files derived from public datasets keep their ODbL and CC BY-SA terms.

Why Tiny Local LLM Projects Matter

Experiments like this are a useful counterweight to the assumption that AI always needs a data center. Serious local LLM work now has a stable llama.cpp foundation, while projects such as TINY show the fundamentals at their most legible. If floppy preservation is your other hobby, the Adafruit Floppsy archiver makes a fitting companion. We have not run this build ourselves, so treat the specifications as the developer's. Find more small-system projects in our mini computers section.

Sources: Hackaday — November 18, 2026; TINY 2.1 on GitHub — accessed November 18, 2026.

More Mini Computers Stories