Z.ai unveils GLM-5.3 open-source AI model with record coding, cybersecurity chops

Key takeaway

• Chinese AI developer Z.ai unveiled GLM-5.3, an open-source large language model that set records on coding and cybersecurity benchmarks. The model performed 50% better than its predecessor on Z.ai's internal coding-agent test and outperformed Anthropic's Claude Mythos 5 on CyberGym, a cybersecurity benchmark measuring code vulnerability detection.

• Z.ai says GLM-5.3 has found over 2,400 vulnerabilities in 269 software projects, with roughly half rated medium severity or higher.

• The model is available now via Z.ai's GLM Coding Plan subscription and will be released on Hugging Face under an open-source license within two weeks.

3 Key Points

FAQ

How did Z.ai train GLM-5.3 to excel at long-horizon coding tasks?

Z.ai used sandbox environments designed to mimic developer workstations and instructed GLM-5.3 to complete complex coding exercises, some of which took days to finish. The company generated sandboxes using specialized AI agents that modeled real-world software projects and created customized programming exercises for each sandbox, with a separate judge agent verifying the challenges were solvable before presenting them to the model.

When will GLM-5.3 be available as open-source?

GLM-5.3 is currently available through Z.ai's GLM Coding Plan subscription service, and the company plans to release the model's weights on Hugging Face under an open-source license within two weeks.

What vulnerabilities has GLM-5.3 found, and how severe are they?

GLM-5.3 has found more than 2,400 vulnerabilities in 269 software projects, with about half rated medium severity or higher. One vulnerability the model identified was in code authored 40 years ago.