Developers have created a tiny μLM AI that runs on a Z80 CPU using only 64KB of RAM.

Unix developer Harry Reed released the Z80-μLM project, a tiny AI model aiming to create a conversational AI model running on a Z80 CPU operating at only 4MHz with just 64KB of RAM. The result is a language model that demonstrates a reasonable level of user understanding while using only 40KB of real-world space, although its responses are limited to specific points.

Reed's model is extremely small, with only 150,000 parameters and using only 2-bit processing. The training process utilizes Quantization-Aware Training (QAT), which trains floating-point parameters alongside 2-bit integers to ensure the model's knowledge remains after quantization. The training process also relies on larger models to generate specialized datasets.

The result is a very small model, such as tinychat, a language model that provides short responses with just a few words but demonstrates understanding of the language being used, or guess, a specialized chatbot that answers 20 questions. Yes or no? The bot uses this question to guess what the bot is thinking about.