A specific 4-bit quantization method that uses a normalized float format to preserve model accuracy while dramatically reducing memory requirements.