Through systematic experiments DeepSeek found the optimal balance between computation and memory with 75% of sparse model ...