Generally HB-MNL estimation (HB, with a multinomial logit formulation of the lower level model fit) is preferred for CBC data. This is what our Lighthouse Studio software does by default when you ask for HB estimation.
For CVA (traditional ratings-based conjoint), our Lighthouse Studio platform also has built-in capabilities for HB-OLS estimation (HB, with an OLS formulation of the lower level model fit). And, this is preferred to to a non-HB individual-level OLS approach.
I don't understand your question about zero-centered diff utility values as a dependent variable. But, I'm guessing that you may be thinking about building a subsequent correlation or regression model where you are using different predictor variables outside the conjoint experiment to predict the preference results of the conjoint analysis. Zero-centered diffs would be preferred for that, since it normalizes the range of utility scores per respondent to be comparable across respondents.
OLS-derived zero-centered diff utilities probably have more noise and overfitting than HB-OLS derived zero-centered diff utilities.