Mixed Low-Bit Quantization For Model Compression With Layer Importance And Gradient Estimations