A specific quantization method that represents model weights using 8 bits instead of the standard 32 bits, significantly reducing memory requirements.