“The DeepSeek model activates 32 out of 256 available experts for a given token.”