[GLM-5.3 by Zhipu Released: Significant Improvements in Coding and Long-Range Task Capabilities, Model Weights to Be Opened in Two Weeks] According to Deep Tide TechFlow, on August 14, Z.ai released GLM-5.3, which is based on the same foundational model as GLM-5.2 but achieves enhanced capabilities through expanded post-training. Official sources state that GLM-5.3 improves coding performance by 50% on the internal Z.ai Code Bench benchmark compared to GLM-5.2 and reaches a leading level among open models in public benchmarks such as Terminal Bench 3.0 and Agents’ Last Exam. In terms of cybersecurity, GLM-5.3 achieved a score of 84.5% in the CyberGym vulnerability detection test and demonstrated significant improvements over its predecessor in exploit chain-related tests such as ExploitBench and ExploitGym. Z.ai announced that the weights for GLM-5.3 will be made available two weeks after release, as the model is currently undergoing security evaluation and reinforcement. Regarding the API, the model no longer supports disabling the "thinking mode" and offers three levels of inference intensity: low, high, and max.