Comment by MattRix

This is exactly it. It’s pretty remarkable that it was trained on over 100 terabytes of images and yet the model has been distilled down to only 4gb.