We present a comparative study of parameters identification of HIV dynamic models for naive patients that are treated with two different HAART (Highly Active Anti-Retroviral Therapy) protocols during a period of 48 weeks. Three HIV models of increasing complexity (in terms of number of state variables and parameters) have been chosen, and for each one the model parameters are computed by solving a nonlinear optimization problem via sequential quadratic programming (SQP). Model parameters are divided into “group dependent”, common to all patients treated with same HAART protocol, and “patient dependent”, specific for each patient, and are estimated in a way that an overall cost function comprising the fitting error of CD4+ concentration and viral load measurements. A preliminary parameter space grid search algorithm is performed in order to find a suitable initial guess for the SQP algorithm. Numerical results indicate that all considered models can give a good matching despite the scarcity of available measurements for each patient, and in this limited situation the minimal model appears to be (slightly) more effective than the other models.