A specific quantization method that converts model weights to 8-bit floating-point numbers using fixed scaling factors, reducing model size while potentially affecting accuracy on complex tasks.