Running CPU-only LLMs without a GPU — what actually works. Best model picks, real speed benchmarks, and a budget dual Xeon server build for 70B models.