Dynamic Watch Beating AI News Flash: The flagship model GLM-5.3 will be open-sourced on August 28 at 10 a.m. Beijing time.

GLM-5.3 was actually launched on the Coding Plan on August 14 and the API was subsequently opened. The model weights will finally be available for download tomorrow, delayed by a full two weeks for a somewhat special reason. Dynamic Watch stated that during the post-training process, the network's cybersecurity capabilities have grown at a rate "beyond expectations." The model not only became better at finding vulnerabilities but also started learning to plan a complete exploit chain across multiple steps. Therefore, the officials first conducted a security assessment and reinforcement before releasing the weights.

In official evaluations, GLM-5.3 scored 84.5% on the vulnerability discovery test CyberGym, slightly higher than Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%. On the more attack-focused ExploitBench, it jumped from 24.4% in GLM-5.2 to 54.4%, but still significantly lagged behind Mythos 5 at 78.0%. Dynamic Watch also stated that the model has discovered 2,436 vulnerabilities in 269 open-source projects, with 1,097 classified as critical or high-risk.

GLM-5.3 inherits the same base model as GLM-5.2, with all these improvements coming from post-training.