At a moment when frontier labs are racing to build trillion-parameter models that fill data centers, one developer has gone the other direction entirely: a working question-answering language model that fits on a single 1980s floppy disk and runs on a CPU old enough to have shipped inside a 486DX.

The project, called Tiny 2.1, was spotted by Hackaday and picked up this week by Futurism and Tom's Hardware. According to its GitHub project page, the entire model weighs in at roughly 380 KB of 8-bit weights — small enough to fit comfortably on one MS-DOS-formatted floppy disk — and needs just 540 KB of memory to run. Tom's Hardware reports that it even speaks its answers through the humble PC speaker. If you enjoy stories about AI breaking expectations, our latest AI developments cover everything from edge inference to trillion-parameter frontier systems.

A Question-Answering Transformer With 367,000 Parameters

Tiny 2.1 is a small transformer architecture with approximately 367,000 parameters — a fraction of the size of the "lightweight" models that typically measure in the low billions. It was written in Turbo C 2.0 and runs on MS-DOS, the operating system that dominated personal computing in the 1980s and early 1990s, targeting machines as old as the Intel 486DX and early Pentium generations.

Its training data is correspondingly tiny: just 3,500 facts. That constraint defines the experience of using it. In a screenshot shared on the project page, the model correctly answers that whales are mammals rather than fish and correctly credits Alexander Graham Bell with inventing the telephone. Asked to compute 17 times 23, however, it replies "96." And asked who the King of France is, it confidently responds "It was Antarce Newton" — with a confidence level of 100 percent.

Don't Expect Claude-Like Performance

As Futurism notes bluntly, the model "makes up answers pretty much all the time," because it simply does not know very much about the world. There is no retrieval layer, no tool use, no web search — just a compact neural network and whatever patterns it absorbed from its 3,500-fact training set.

What makes the project notable is less what the model can do than what its existence demonstrates. Modern LLM inference is associated with gigabytes of memory, GPU acceleration, and kilowatts of power. Tiny 2.1 instead proves the concept that a transformer can produce structured, question-and-answer interaction on hardware with no floating-point unit to speak of, no hard drive requirements beyond a floppy, and an audio output path as primitive as a motherboard speaker.

AI Built With AI

There is also a meta twist. According to the project page, the developer did not write every line by hand: they used Anthropic's Fable 5.2 to help write the model's code and Opus 5.5 to support the training process. In other words, a model small enough to fit on a floppy disk was itself authored and trained with the assistance of frontier-scale AI systems — a tiny descendant of very large parents.

That workflow mirrors a broader trend in AI engineering, where large models increasingly generate, debug, and optimize the systems that run on much smaller hardware. The same pattern shows up in modern edge-AI toolchains, where cloud-scale models help compress and quantize networks for phones, microcontrollers, and embedded devices.

Why Tiny Models Matter

Tiny 2.1 will not threaten any commercial chatbot. But projects like it serve as a useful reminder of how far model compression has come — and how much bloat is choice rather than necessity. Retro-computing enthusiasts have spent years demonstrating that decades-old hardware remains capable of surprising work, from browser-capable Amigas to networked 8-bit machines. A language model joining that lineage is a first, and a playful one.

The engineering constraints that shaped Tiny 2.1 are also a crash course in efficiency. An 8-bit weight format, a vocabulary and architecture sized to the task, and an output path designed around hardware limitations add up to a system where every kilobyte is accounted for. Contrast that with mainstream deployments, where quantization, pruning, and distillation are retroactive fixes applied to models built at vastly larger scale — and where a single chat session can consume more memory than the entire Tiny 2.1 runtime needs, weights included.

It also lands at a time of real environmental scrutiny around AI's power appetite, with researchers and regulators increasingly focused on the energy cost of large-scale training and inference, and data center siting fights playing out across the United States. A model that runs on 540 KB of memory and a 1980s power supply is not a solution to that problem — but it is a vivid illustration of the other end of the spectrum: inference with a footprint measured in watts, not megawatts. Readers tracking how far the field stretches, from floppy disks to gigawatt campuses, can find both ends in our breaking AI news.

For now, Tiny 2.1's maker can claim a genuinely rare achievement: a neural network you can install from the same medium you'd use to save a word processor document, on a machine that predates the web. It hallucinates constantly, misquotes the identity of French royalty with total confidence, and cannot multiply two-digit numbers reliably. It also runs, entirely, on a floppy disk.

Stay Ahead of AI

From floppy-disk models to frontier labs, we track it all. Read more AI news →