In the latest version, it only needs 64 review logs to full-optimize the model