Chinese startup Z ai has officially released a new GLM-5 language model.3 with improved programming, agent-based scripting, and vulnerability detection capabilities. According to the developers, there is a significant increase in performance compared to the previous version of GLM-5.2 was obtained solely through additional post-training without replacing the base model. The open weights of the new neural network are planned to be published two weeks after the completion of the security assessment and additional protection enhancement. The GLM-5 model is based on the internal Code Bench benchmark for evaluating coding skills.3 improved the result of its predecessor by fifty percent.

In the official release table, Z ai provided a detailed comparison of the results for a number of specialized tests. On Terminal Bench 3.0, the new model scored twenty-eight and three tenths points against four and six tenths for the GLM-5.2. On DeepSWE v1.1, the indicator was sixty-six and nine tenths versus forty-six and two tenths. The Agents' Last Exam showed a score of twenty-eight and five tenths against twenty-three and eight tenths points. On the CyberGym benchmark, the model reached eighty–four and a half percent versus seventy-seven and two tenths, and on ExploitBench, it reached fifty-four and four tenths percent versus twenty-four and four tenths percent for the previous version. Separately, the developers highlighted the test results for ExploitGym, where the model successfully completed one hundred and five tasks with a two-hour time budget compared to twenty-nine for the GLM-5.2. With an increase in the time budget to six hours, the indicator increased to one hundred and thirty tasks instead of thirty-nine in the previous version.

The release focuses on testing on real-world code bases in collaboration with several cybersecurity teams in China. After careful verification and deduplication, the system identified two thousand four hundred and thirty-six vulnerabilities in two hundred and sixty-nine projects, including one thousand ninety-seven problems of medium and high criticality. Fifty-three finds have been publicly disclosed, while another two thousand three hundred and eighty-three remain under embargo. The company estimated the average lifetime of such vulnerabilities at twenty-six point six years, and dated the oldest of the detected problems to nineteen eighty-one. To track the discovery process, a special public registry, the Z ai Security Disclosure Ledger, was launched, which records the status, affected projects, and assigned criticality level. At the same time, the company updated its product line by deploying GLM-5.3 for subscribers of the GLM Coding Plan as the main engine for agent-based programming and long-running tasks. All requests to the old models are now automatically routed to the new version.