3 ms·

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