An 80GB H100, and only 3B of the model could keep up with conversation

Even with an H100 (about 5 million JPY), only a model using 3B per word could be employed for real-time conversation. Reflecting on the challenges of MoE and voice dialogue based solely on text judgment results.