Zhipu Tangjie: GLM-5.3 expanded and trained on the same base
According to Tang Jie of Zhipu AI, GLM-5.3 and GLM-5.2 use the same base, architecture, total parameters and activation parameters, and are mainly trained through a long-term environment and reinforcement learning expansion over a month.
Tang Jie of Zhipu AI stated that GLM-5.3 and GLM-5.2 use the same base, architecture, total number of parameters and number of activation parameters, with the main changes coming from expanded training.
According to their explanation, the team spent a month expanding the long-range environment and reinforcement learning, using GLM-5.3 as a controlled experiment for post-training the Scaling Law. This statement indicates that capability improvement does not solely depend on increasing parameters, but it represents the R&D team's firsthand experimental interpretation and cannot replace independent benchmark replication.
source:Tang Jie's public statement。