Compared to traditional Artificial Neural Network (ANN), Spiking Neural Network (SNN) has garnered widespread academic interest for its intrinsic ability to transmit information in a more biological-inspired and energy-efficient manner. However, despite previous efforts to optimize the learning gradients and model structure of SNNs through various methods, SNNs still lag behind ANNs in terms of performance to some extent. The recently proposed multi-threshold model provides more possibilities for further enhancing the learning capability of SNNs. In this paper, we rigorously analyze the relationship among the multi-threshold model, vanilla spiking model and quantized ANNs from a mathematical perspective, then propose a novel LM-HT model, which is an equidistant multi-hierarchical model that can dynamically regulate the global input current and membrane potential leakage on the time dimension. In addition, we note that the direct training algorithm based on the LM-HT model can seamlessly integrate with the traditional ANN-SNN Conversion framework. This novel hybrid learning framework can effectively improve the relatively poor performance of converted SNNs under low time latency. Extensive experimental results have demonstrated that our LM-HT model can significantly outperform previous state-of-the-art works on various types of datasets, which promote SNNs to achieve a brand-new level of performance comparable to quantized ANNs.