4 ms·

For those that prefer text: it seems they use a weaker but faster model for the "predicted output" / speculation. Pretty smart.