Every model starts from a slightly different guess about right now, because we do not measure the atmosphere everywhere. Small differences at hour zero become large differences at hour seventy-two. That is not a bug, that is the atmosphere.
Models also differ in physics. How one handles convection, or the boundary layer, or snow ratios, will bias it in known directions. Knowing those biases is most of the skill.
Ensembles exist to show you the spread. A tight ensemble is a confident forecast. A wide one is the model telling you it does not know, and you should pass that honesty along instead of averaging it into a single false number.
When the models fight, say so. 'The models are fighting' is a real forecast state, not a cop-out.