These models are quantized in mixed precision that allows them to have a smaller footprint than fp8, but still high quality.