Fascinating stuff! I'm not familiar with these new reasoning based models, I wonder how much of the reasoning ability comes from the weights vs the architecture of the model itself vs the system prompt.