Community finds MiniMax H3 effectively 20.1B params, not 33.1B

Reddit analysis of MiniMax H3 says the 33.1B-parameter model effectively uses 20.1B because AdaLN projections were built to accept 2688 inputs but receive only a timestep value during inference. The community sees the smaller effective size as a positive for efficiency.