Specific Emitter Identification (SEI) distinguishes radiation sources by extracting unique signal features. However, when deploying such models, the large number of parameters often leads to excessive memory usage, hindering practical application. To address this issue, we propose a dynamic adaptive quadratic quantization scaling algorithm, which treats the scaling factor as a trainable parameter and dynamically optimizes quantization boundaries during training. Additionally, we employ a quantile Huber loss function to reduce the sensitivity of quantized weights to quantization errors, enhancing model robustness. A regularization term is also introduced to facilitate optimizer convergence. Experimental results demonstrate that our approach achieves low-bit-width model compression with an accuracy loss of less than 2%, closely matching the performance of the original model. This method significantly reduces memory consumption, compresses model size, and maintains recognition accuracy, enabling efficient inference on resource-constrained devices.