This is crazy! More than 90% of the Hy3 model parameters stay idle during inference. At first, that sounds inefficient. Modern AI models don't try to use everything they know all at once. They decide which experts should solve each token. Instead of asking the entire model to work on every token, it routes the work to a small set of specialized experts. You get much of the capability of a large model without paying the full computational cost every time. I spent some time experimenting…