Z80-μLM is a 2-bit quantized language model small enough to run on an 8-bit Z80 processor (github.com)

Z80-μLM introduces a groundbreaking 'conversational AI' model that operates on an 8-bit Z80 processor with just 64KB of RAM, showcasing quantization-aware training (QAT) within a mere 40KB .COM binary. This innovative project was sparked by the pursuit of creating a compact and personality-driven AI that could be easily self-hosted, and it reveals a fascinating blend of nostalgia and modern AI capabilities. Although it won’t pass the Turing test, its simple yet charming interactions, such as a casual chatbot and a 20 Questions game, underscore its potential for engaging conversational experiences. For the AI/ML community, Z80-μLM is significant as it explores the limits of hardware-constrained AI, challenging the perception that advanced machine learning requires substantial resources. The technical implementation includes trigram hash encoding for typo tolerance, integer-only arithmetic for stability, and an autoregressive text generation mechanism—all custom-tailored to the limitations of the vintage Z80 CPU. This project not only reminds developers of the ingenuity possible within constrained environments but also invites experimentation in generating unique machine learning models that prioritize efficiency and creativity.