A practical breakdown of which large language models run on CPU-only hardware, how fast they go, what quantization and inference engines to use, and where CPU inference fits in a real production stack.