Z.ai Ships GLM-5.3 Without Retraining the Base Model: Better at Complex Coding and Long-Horizon Tasks

Z.ai released GLM-5.3, a new version of its language model that uses the same underlying base model as its previous version but achieves improvements through additional post-training on more tasks and environments. The model shows substantial gains in coding tasks, with one benchmark jumping from 4.6 to 28.3, and unexpectedly strong performance in cybersecurity applications like vulnerability discovery, where it outperformed some competing models. The weights are not yet publicly available but will be released in approximately two weeks after safety evaluation, though the model is already accessible through Z.ai's API and coding plan for some users. This matters because the improvements demonstrate that scaling post-training alone, without retraining the base model, can produce significant capability gains in complex tasks like long-horizon coding and security analysis.