TechFlow news, on August 19, Zhipu AI founder Tang Jie stated in a post that large model scaling should not focus solely on parameter count, but also simultaneously consider data scale, computational resource allocation, inference costs, and actual operating conditions. The article reviewed the evolution from Kaplan's Scaling Laws to Chinchilla's compute-optimal theory, pointing out that the early route of "parameter growth faster than data growth" once led to resource misallocation in some super-large models, while subsequent research indicates that model parameters and training data should grow in a more balanced manner.
The article further pointed out that as models are invoked frequently, inference costs have become the core of model lifecycle costs, and the optimal strategy is shifting towards smaller models and longer training; meanwhile, in Mixture of Experts models, total parameters affect knowledge capacity more, activated parameters and effective depth affect inference capability more, and the single "training tokens per parameter" ratio is no longer applicable.
Regarding GLM-5.3, the article stated that it uses the same base model, architecture, total parameters, and activated parameters as GLM-5.2, achieving significant improvements through one month of post-training involving scaled long-context environments and reinforcement learning. The article believes that parameter scale, pretraining data, forward computation, and post-training are all independently adjustable scaling variables; this experiment indicates that post-training still has significant room for improvement, and scaling has not yet ended.





