What they did
The poster wanted the small TINY 2.1 language model, with its original weights, running on an unmodified Amiga 500: a 7 MHz Motorola 68000 with no floating point unit and 1 MB of memory. They reused the workflow from an earlier port of a TinyStories model to the same machine and had Claude Code with Opus 5.5 write the C inference engine, the 68000 assembly and a GUI, all in integer math. Each change was checked bit for bit against an integer reference written in Python, then in an emulated 68000, then in the WinUAE emulator, and only then on the real computer. They also had it port the Amiga CD32 intro for the demo.
How it went
In exact mode the Amiga gives the same answers as the original model on all 44 test questions. It produces 0.2 to 0.4 letters per second, with a KV cache and a few other optimisations to get there; the demo video is sped up in marked places. The code is published as open source.
Worth knowing
The verification chain is what made it work on a machine the agent could not run itself: every step had a reference to compare against. You still need the TINY.BIN weights file from the original model's repo. The poster did not say how many sessions it took or what it cost.