Compared to traditional models, Meituan's large model shows significant advantages in training and inference speed, achieving a single-user inference speed of over 100 tokens per second on H800, with ...