MiniMax H3 effectively runs at 20.1B params despite 33.1B architecture

Community testers found MiniMax H3's AdaLN projections were built to accept 2688 inputs but receive only 1 (a timestep value), cutting effective parameters to 20.1B of its 33.1B. The discoverers call it "a VERY good thing" — the model is far more efficient than its declared size suggests.