On August 14, Zhipu (02513.HK) made headlines by announcing the launch of its latest GLM-5.3 model. Constructed upon the same foundational architecture as its predecessor, the GLM-5.2, the enhanced performance of this new model is attributed to meticulous post-training refinements. When pitted against GLM-5.2, the GLM-5.3 model showcases remarkable advancements in tackling intricate coding challenges and managing extended-duration tasks, solidifying its position as the most formidable open-source weight model currently available.

Internal benchmark tests conducted by Z.ai on coding proficiency reveal that the GLM-5.3 model has achieved a significant 50% performance boost over the GLM-5.2. Moreover, in large-scale deployment scenarios, the network capabilities of GLM-5.3 have surpassed initial projections. On the CyberGym platform, the GLM-5.3 model stands out with its leading-edge vulnerability detection prowess, particularly excelling in the latter phases of exploitation chains. Its performance in exploitation benchmark assessments has more than doubled compared to that of the GLM-5.2.

Following the completion of rigorous security evaluations and reinforcement measures, Zhipu is set to release the model weights to the public two weeks subsequent to the announcement.