Microgpt in pure C hits 10M tps on Apple m5

MicroGPT-C is a lightweight GPT implementation written in pure C with no external dependencies, achieving 10.1 million tokens per second on Apple's M5 Pro. The character-level transformer model can train on ~32,000 names in seconds and includes forward pass, backpropagation, Adam optimization, and sampling all in a single C file. It successfully generalizes on unseen data, outperforming larger trigram models while supporting multiple platforms and CPU architectures including ARM64 with NEON and x86-64 with AVX2.