An analytical modeling method for heterojunction bipolar transistor (HBT) is proposed in this paper. The new neuro-space mapping (Neuro-SM) model applied to DC, small signals and large signals simultaneously consists of two mapping networks, which provide the additional degrees of freedom. Sensitivity analysis expressions are derived to accelerate the training process. When the nonlinearity of device is high, or the response of the model is complex, the weights in the proposed model are automatically adjusted to address the accuracy limitations. The proposed modeling method is verified by measured HBT examples in DC, small signals and large signals Harmonic Balance (HB) simulation. The modeling experiments of the measured HBT demonstrate that the errors of the proposed Neuro-SM model are less than 2% by matching combined DC, small-signal S-parameters and large-signal HB data, which are less than the errors of the traditional Neuro-SM model and the coarse model. The proposed analytic Neuro-SM model fits the response of the fine model well.