Chinese technology firm Z.ai has announced the launch of its open-weight latest large language model (LLM) GLM-5.3, which it claims is competitive with Western frontier models and better than them at cybersecurity tasks.
The lab said GLM-5.3 is “the most capable open-weights model for coding” and pushes the frontier of cyber vulnerability discovery.
In the CyberGym benchmark, which tests a model’s ability to reproduce vulnerabilities in an unpatched codebase, GLM-5.3 scored 84.5 per cent. This is more than the 83.8 per cent achieved by Anthropic’s Claude Fable 5 or the 83.6 per cent by OpenAI’s GPT-5.6 Sol.
Z.ai said that it specifically trained the model to improve at cyber tasks but was surprised to discover the post-training allowed GLM-5.3 to build multi-stage exploits. It added that in tests with Chinese security teams, GLM-5.3 identified 2,436 vulnerabilities across 269 projects, of which approximately 1,097 were medium-to-high severity.
GLM-5.3 also competes directly with other leading Chinese models, including Moonshot AI’s flagship Kimi K3 and DeepSeek V4 Pro.
In the frontier coding benchmark Terminal Bench 3.0, the model scored 28.3, a 615 per cent improvement on GLM-5.2 that puts the model ahead of Kimi K3 for code generation but behind the latest models from Anthropic and OpenAI.
Z.ai said GLM-5.3 was trained on the same 743-billion foundation model as its predecessor, with every benchmark gain having been achieved through post-training alone. Sometimes called “alignment”, this is the process by which the raw outputs of an LLM are optimised through supervised fine-tuning and reinforcement learning to cement positive “behaviours” in the model and ensure its outputs are high quality.
At present, GLM-5.3 is available via Z.ai’s coding agent ZCode, as well as its top subscription tier GLM Coding Plan. It said it will make the model available via API and release its open weights in two weeks, “following rigorous safety evaluations”.
