Zhipu has announced that GLM-5 represents a next-generation foundational model that is propelling the evolution of programming paradigms, shifting from 'VibeCoding' to 'AgenticEngineering'. Leveraging the ARC (Advanced Reasoning and Comprehension) capabilities of its predecessor, GLM-4.5, this model significantly cuts down on inference expenses through the implementation of sparse attention mechanisms, all the while preserving its ability to process long contexts. Moreover, GLM-5 has established a cutting-edge asynchronous reinforcement learning framework, boosting the efficiency of post-training iterations. It also introduces an innovative asynchronous agent reinforcement learning algorithm, which further refines the outcomes of reinforcement learning. Thanks to these technological advancements, GLM-5 has attained state-of-the-art (SOTA) performance in open benchmark evaluations and outperformed all open-source baselines in real-world programming tasks as well as end-to-end software engineering endeavors.