Z.ai says GLM 5.3 was built by post-training GLM 5.2 without new pretraining

Z.ai's GLM 5.3 was built entirely by post-training its predecessor, GLM 5.2, without pretraining a new base model, according to a SemiAnalysis post on X. The post quotes an unnamed speaker who says post-training is far more GPU-efficient than pretraining, and that GLM 5.3's cyber capabilities are dramatically better than GLM 5.2's, all from post-training. The report does not include independent benchmark results or Z.ai's own statement on the training method.