The performance of large models has been improved, and the freezing of the model should not happen anymore.
