DeepSeek Pro's 1.6T total parameters with 49B activated parameters result in inference costs comparable to a roughly 49B dense model, while expanding knowledge capacity by approximately 32 times.