gemma4.c runs Gemma 4 E2B in 700 lines of pure C at real time on a Ryzen 7 7700.

One file handles tokenizer, transformer, KV cache, sampling and CPU kernels with no external library doing the interesting parts, producing about a 5.0 GB model file (GitHub). Weights are int8 with FP16 scales, linear-layer inputs dynamically quantized to int8, other activations float32, validated against the Hugging Face reference running Google's unquantized QAT checkpoint in BF16. The author confirms the demo is real time, CPU only. Read it if you want to know what a competent inference runtime is actually doing.