Probably quantizing or using base weights and this project https://github.com/ggerganov/llama.cpp on a CPU machine with AVX512 instructions.