A model held at eight or four bits per weight instead of sixteen takes far less memory and runs on far cheaper hardware. For a studio running anything locally, this is the difference between one workstation and a rack.
The quality cost is real but non-linear. Moderate quantisation is often visually indistinguishable; aggressive quantisation shows up first in exactly the places that matter, meaning fine detail, small text and subtle colour.
If you run local inference, record the quantisation level alongside the model version. Two operators running the same model at different precisions will produce a set that does not match, and nobody will guess why.
Does quantisation change the output?
Yes, subtly, and first in fine detail, small text and colour subtlety. Moderate levels are often invisible; aggressive levels are not.
Why record the quantisation level?
Because two operators running the same model at different precisions produce sets that do not match, and it is close to undiagnosable unless it was written down.