AI-Assisted Coders Run Windows on iPhone Duo, Build a Floppy Disk-Based AI Chatbot for MS-DOS
2 43The blog XDA Developers reports that a Claude enthusiast used the AI assistant to make a language model that runs off a floppy disk. "And you can stick it on a DOS computer to boot."
The language models "runs at a respectable speed, given that it's running off a CPU that was designed decades before LLMs were even a thing." Over on the ClaudeAI subreddit, user wollymonjak showed off this really impressive feat... [I]t's not going to be a particularly powerful AI; in fact, wollymonjak notes that if you ask something too complex or too long, it starts spitting out nonsense. But who cares about that when you're using a language model on DOS?
The same blog also notes that "the iPhone Duo hasn't been out for too long, and someone has already gotten Windows 11 on it." As spotted by TechSpot, this cool feat was performed by Ammaar Reshi over on X. Their project is called WindowsiOS, and its goal is to get Windows 11 running on an iPhone Duo without jailbreaking. Usually, when we cover a project coded up by AI, Claude is the one that called the shots; this time, Reshi opted to use Gemini 4 Carbon, Google's flagship model.
2 comments
Insanely small (Score: 5, Interesting)
by Rei ( 128717 ) on Monday November 16, 2026 @07:53AM (#66426622)
A "tiny" LLM is generally considered to be about 1B params or less. This one is 366k params.
Of course, it seems to be immensely narrowly trained: "Trained on about 3,500 facts, each asked in many wordings Outside those it makes things up confidently"
What would be more interesting, IMHO, would be to try to train a RAG model. Something that is trained on *no* facts, merely on converting your statements into search terms, and turning search results back into answers. I wonder how small you could get that? I've often wondered how far back we technically could have run an actually-useful LLM if we were to have had the weights and inference engine.
Re:Insanely small (Score: 5, Interesting)
by Rei ( 128717 ) on Monday November 16, 2026 @08:17AM (#66426662)
ED: Looking into it, the state of the art for actually-useful RAG is still about 60-100M params, so still way off from what you can fit on a floppy disk. Quantized, you're looking at maybe 25 floppy disks, 30-50MB at 4-bit compressed.
For running it comfortably, a '96-98 era Pentium MMX / Pentium II fully-in-ram with 64-128MB ram and a 1-4GB hard drive should be easy and run at a workable (go brew a cup of coffee ;) ) 1 tok/s out.
For at least being able to fit it in RAM, you're looking at the 1987-1990 era - a Mac II with the FDHD could support up to 64 MB of RAM, and a 386DX could handle 32-64MB on some motherboards. It would have cost tens of thousands of dollars and taken like half an hour per answer.
If you're really damned patient and willing to stream from disk, we can go back to 1984-1986 (IBM PC AT / XT + 40MB HDD). It could just keep the current hidden state and a tiny 640k weight buffer in RAM and stream the model layer by layer. Prefilling a tiny 300 token passage would take like 2 days, and a 40 token answer would take like 5 hours. On a 4.77 MHz 8088 PC, you're talking maybe 1-2 weeks per query.
And if you're a masochist... technically you could stream straight off of manually-swapped floppy disks! Then you could run it on a 1977-1982 Apple IIe or Commodore 64 ;) It'd take ~200x 5 1/4" disks. Of course, merely the stage of generating a 30 word answer would take around a month and six thousand floppy swaps. ;) Once every ~7 minutes.