tinygrad is an ultra-minimal deep learning framework that implements every modern neural network layer using only 12 primitive operations, making it remarkably approachable and hackable. Despite its microscopic codebase measured in thousands of lines, it powers production neural networks inside comma.ai's openpilot and can run Meta's LLaMA and Stable Diffusion with no dependencies beyond NumPy and a C compiler. The tinybox hardware product, built on tinygrad, now makes it possible to run powerful neural networks entirely offline.