does alphafold 2 explicitly model co-evolution?

I am looking for some clarification on the role of co-evolution in alphafold 2.

From my understanding:
Extracting information about which residues have co-evolved (and, as such, are likely to be in contact) has historically been an important processing step. However, I was under the impression that this is no longer done explicitly in Alphafold 2. Rather, the model learns itself how to process the raw MSA. It COULD learn to reconstruct co-evolution patterns, among many other things, but we don't explicitly tell it what to do.

I am asking this because resources are not very clear about this. For example, the Nobel price overview of the model: https://www.nobelprize.org/uploads/2024/10/fig2_ke_en_24.pdf
talks a lot about co-evolution.
But is this our assumption about what the model learns from the MSA? For all we know, there are different patterns in the data the model relies on much more.

I looked through the code and found no specific MSA pre-processing like that, and co-evolution is never mentioned in the Alphafold2 paper.

thank you!