It's a tiny 6M parameter text to image model, and it can make recognizable images at 64x64. The 6M accounts for everything, there's no oversized text encoder or VAE. Check out the model card for samples, a detailed explanation about it's architecture and training, and a quick start!
