Z.ai has released an update to its GLM-5.3 model, which builds upon the existing 743B base model without retraining. The update focuses on scaling post-training processes, resulting in improved performance on complex coding and long-horizon tasks. According to reports, the model has shown significant gains in various benchmarks, including Terminal-Bench 3.0 and DeepSWE v1.1. The update also demonstrates advancements in cybersecurity, with notable improvements in CyberGym and ExploitBench. The new weights for the model are expected to be available in about two weeks. This development is significant as it highlights the potential for incremental improvements in AI models without the need for complete retraining.