Classic diffusion learns a winding route from noise to image and has to take many small steps along it. Flow matching learns a straighter one, so fewer steps get you to the same place.
The visible effect is that recent models produce usable output at step counts that would have been unusable two years ago, which lowers cost per render and makes proxy passes genuinely cheap.
It is not a control you operate. It is a reason the defaults on a current model are better than the defaults you learned on an older one, which is worth knowing before you copy old settings forward.
Do I need to change my settings for a flow-matching model?
Usually yes, downward. Step counts and guidance values carried over from older diffusion models are often too high and will cost you money and sometimes quality.
Is this the same as distillation?
No. Distillation trains a fast model to imitate a slow one. Flow matching changes how the path from noise to image is learned in the first place.