August 15 news.ZhipuYESTERDAY GLM-5.3 WAS PUBLISHED。New ModelFollowing the GLM-5.2 base, there was no re-training, and the increase in capacity came mainly from post-training in longer, larger and more diverse environments. Officially, GLM-5.3 is online with Zcode, AutoClaw, GLM Coding Plan and multiple coding platforms, and API will then be open, and model weights are scheduled to open within two weeks of the completion of security assessments and enhancements。
Programming: Terminal-Bench 3.0 score from GM-5.2 to 28.3, DeepSWE v1.1 from 46.2 to 66.9 and Ages 'Last Exam from 23.8 to 28.5。
Code execution efficiency: Self-building Z.ai Code Bench shows that the GLM-5.3 accuracy of the High slot is 31.4%, with an average output of approximately 50,000 tokens per job; by contrast, Claude Opus 4.8 has a maximum level of 29.5%, with an average output of about 120,000 tokens。
Net security: GLM-5.3 score in CyberGym 84.5%, slightly higher than 83.8% of Mythos 5 and 83.6% of GPT-5.6 Sol; ExpluitBench score 54.4%, still below two closed-source models。
Security assessment: 2436 gaps, of which 1,097 are high- and medium-risk, involving 269 projects, have been identified and removed by the intellectual and higher education and security teams; the gap has been repaired and will enter the public disclosure book, and only Hashi values are shown for projects still being coordinated for disclosure。
At the same time, the "open source shield" program was launched to provide an ongoing security audit, model scale and ZCode code audit portal to open source projects. According to the company, the two-week extension of the weighting was intended to limit the potential for attack and to preserve defensive use。