Am I wrong to assume that large MoE models are likely the point anyway? Seems like this card would love DeepSeek V4.1 Flash.