The 405 billion parameter Llama 3.1 model produces only one token every several seconds of real wall-clock time on a high-end workstation, making response generation effectively unusable for interactive dialogue.