Four Hourglasses

Four autoencoder designs are proposed for the same data: inputs of $64$ numbers, rebuilt as $64$ numbers. Each is listed layer by layer from input to output, along with the layer its designer calls the code.

design	layer widths, input to output	layer called the code
A	$64 \to 16 \to 64 \to 16 \to 64$	the second $16$
B	$64 \to 128 \to 8 \to 128 \to 64$	the $8$
C	$64 \to 2 \to 64$ , and the decoder is also handed the original $64$ input numbers directly, alongside the code	the $2$
D	$64 \to 6 \to 12 \to 64$	the $12$

Apart from the extra route in design C, every layer is computed only from the layer immediately before it.

Which design forces the tightest squeeze on what the decoder has to rebuild the input from?

Four Hourglasses

Hints