This is the sort of thinking one brings from Latent Class, where the goal is to maximize the LL.
HB does not have as a goal to maximize the LL of the fit to respondents' choices. In fact, it purposefully gives up a lot of fit at the individual level to the individual's choices in favor of making sure the respondents seem to be drawn from higher density areas of a multivariate normal distribution, according to population means and covariances.
If your two groups have quite different preferences, then the compromise toward population means and variances will yield higher individual-level fit to respondents' choices than when using a pooled model. So, it's probably the case that you will obtain higher individual-level fit on average across respondents when using a segmented model. However, segmented models also have lower sample size within each HB model to stabilize the means and especially the covariances. So, the better individual-level fit isn't necessarily an indication of better quality. There may be some overfitting, for example.