In this paper, an enhanced dynamic Neuro-space mapping (Neuro-SM) method is proposed with emphasis on transistor modeling. By modifying the dynamic voltage relationships in an existing nonlinear model, the proposed Neuro-SM produces a new and more accurate model than the nonlinear model as well as the static Neuro-SM. Compared to the existing dynamic Neuro-SM, a new sensitivity analysis technique is derived to speed up the training of the proposed model with dc, small- and large-signal data. The validity and efficiency of the proposed Neuro-SM method are demonstrated by modeling examples of a GaAs high-electron-mobility transistor (HEMT). Suitable value of time delay parameter which is equal to one divided by 3 or 5 times of the largest frequency considered in simulation is suggested and demonstrated by the modeling example.