GuppyLM Shows a 9M-Parameter Language Model Can Be Trained From Scratch in Minutes

GuppyLM, a tiny open-source language model built to role-play as a fish, is drawing developer attention for showing that a working transformer can be trained from scratch in about five minutes on a free Colab T4 GPU. The project is not a frontier model, but its appeal is precisely that it exposes the full stack: synthetic data generation, tokenizer training, model architecture, training loop and local inference.