GuppyLM drops as 9M parameter educational LLM

GuppyLM is the "Minix for LLMs," offering a readable entry point for developers overwhelmed by the complexity of modern foundation models. It successfully trades state-of-the-art performance for extreme educational clarity.

• –Pure vanilla PyTorch implementation (~130 lines) makes the transformer architecture actually readable in one sitting.
• –The 9M parameter size and 128-token context window are perfectly tuned for educational experimentation on commodity hardware.
• –By intentionally avoiding modern optimizations like RoPE or GQA, it lowers the barrier to entry for understanding the core mechanics of attention.
• –The specialized "fish" persona demonstrates the power of dataset curation and behavior training on a micro-scale.
• –At ~10MB, the quantized model is small enough to deploy via WASM for browser-based inference.