Zhipu AI releases GLM-5.3 to rival Anthropic in coding and security

Zhipu AI (Z. ai) has released GLM-5.3, a new open-weight large language model designed to compete with leading Western models like Anthropic’s Mythos 5. The model utilizes a 743B parameter Mixture of Experts (MoE) architecture, with approximately 40B active parameters, and features a 1-million-token context window.

The update focuses on significant improvements in coding and cybersecurity through post-training rather than changes to the underlying base model. Zhipu AI reports a 50% increase in coding performance compared to the previous GLM-5.2 version. In cybersecurity benchmarks, the model achieved an 84.5% score on the CyberGym test, slightly surpassing Mythos 5’s 83.8% in vulnerability detection.

Despite these gains in detection, the model still trails Western competitors in offensive capabilities. In the Exploit Bench test, which measures the ability to convert vulnerabilities into functional attacks, Mythos 5 scored 78.0% compared to GLM-5.3’s 54.4%. The model is available via the GLM Coding Plan and ZCode, with open weights expected to be released following safety evaluations.