Comment by slavboj

It is not "slow and expensive", although it could be "or". You can get 3 tokens / second running on DDR4 memory on a two generation old workstation system that costs ~1K, via llama.cpp .