Show HN: Z80-μLM, a 'Conversational AI' That Fits in 40KB
3 months ago
- #AI
- #Z80
- #retrocomputing
- Z80-μLM is a conversational AI designed for Z80 processors with 64KB RAM, using quantization-aware training (QAT).
- The project explores minimalistic AI with personality, resulting in a 40KB .com binary running on a 4MHz 1976 processor.
- Features include trigram hash encoding for typo-tolerant input, 2-bit weight quantization, and 16-bit integer inference.
- Includes a chatbot and a 20 Questions game, demonstrating terse, personality-driven responses.
- Training tools provided for generating data with LLMs (Ollama/Claude API) and balancing class distributions.
- Technical highlights: no floating point, autoregressive generation, and tight Z80-native arithmetic loops.
- Despite limitations (e.g., no deep context tracking), it offers a unique, constrained interaction mode.
- License options: MIT or Apache-2.0.