Chinese AI lab Z.ai has shipped GLM-5.3, an open-source large language model that set records on several coding and security benchmarks.
The model keeps the architecture of GLM-5.2, a mixture-of-experts design with 753 billion parameters and a 1 million token context window, but adds a heavier post-training pass. Z.ai says GLM-5.3 scored highest among open-source models on Terminal Bench 3.0, which tests command-line scripting, and performed 50% better than GLM-5.2 on its internal coding-agent benchmark.
Security is the standout result. GLM-5.3 scored 84.5% on CyberGym, a benchmark for finding code vulnerabilities, beating Anthropic’s Mythos 5 on that test. Anthropic’s model still leads in turning discovered flaws into working exploits, and it won on two other cybersecurity benchmarks.
In testing, the model has flagged more than 2,400 flaws in 269 software projects, roughly half of them medium severity or worse. One of the discoveries sits in code written four decades ago.
The model’s weights are set to land on Hugging Face under an open license inside two weeks. Until then, developers can reach it through Z.ai’s GLM Coding Plan subscription.
