AI Model Runs on 64 kB of RAM and a Zilog Z80 CPU

In 1976, the Zilog Z80 CPU was launched, and it’s driven most everyday devices, including calculators, arcade systems, and home PCs. Recently, HarryR (via github) developed the Z80-μLM, an AI tool running on a Zilog Z80 CPU with 64 kB of RAM. While this is a notable achievement, it doesn’t meet the standard of human-like intelligence.

HarryR wrote in the Readme, "Z80-μLM is a 'conversational AI' that generates short character-by-character sequences, with quantization-aware training (QAT) to run on a Z80 processor with 64kb of RAM." The developer wants to explore how AI can be minimized, even with a personality. His project aims to determine if the AI tool can be trained and fine-tuned under extreme constraints. This model achieves that in 40 kB with a chat UI, weights, and inference.

According to HarryR, the Z80AI project features:

• Trigram hash encoding: Input text is hashed into 128 buckets — typo-tolerant, word-order invariant.
• 2-bit weight quantization: Each weight is {-2, -1, 0, +1}, packed 4 per byte.
• 16-bit integer inference: All math uses Z80-native 16-bit signed arithmetic.
• ~40KB .COM file: Fits in CP/M's Transient Program Area (TPA).
• Autoregressive generation: Outputs text character-by-character.
• No floating point: Everything is integer math with fixed-point scaling.
• Interactive chat mode: Just run CHAT with no arguments.

There are two pre-built examples in this project, Tinychat and Guess. Tinychat is a minimalist conversational agent that outputs a short answer in response to greetings and questions. Guess is a game with 20 questions, and the model has a hidden answer that the user must uncover through dialogue.

Both are distributed as precompiled binaries for CP/M systems and the Sinclair ZX Spectrum. On CP/M, the tool uses buildz80com.py to generate standard .COM files that can be run directly. Meanwhile, the ZX Spectrum generates two .TAP files that load on an emulator or original hardware. It uses ZX Spectrum ROM routines for input and output, and the memory is optimized for 48K systems.

It also has short, nuanced responses:

• OK - acknowledged, neutral
• WHY? - questioning your premise
• R U? - casting existential doubt
• MAYBE - genuine uncertainty
• AM I? - reflecting the question back

"It's a different mode of interaction. The terse responses force you to infer meaning from context or ask probing direct yes/no questions to see if it understands or not," said HarryR.