[johndough]

Why not? The model is 2.8T parameters with native MXFP4, which is 1400GB.

[alightsoul]

I would figure it needs a lot of kV cache and context size which can be much bigger than the model itself