Z.ai has launched GLM-5.3, claiming it as the leading open-weights coding model with significant improvements in efficiency and cybersecurity.
Z.ai Introduces Cutting-Edge GLM-5.3 Model
Z.ai, the AI lab previously known as Zhipu AI, has launched its latest coding model, GLM-5.3, claiming it to be the most sophisticated open-weights coding model to date. This new release is part of the company’s GLM Coding Plan and ZCode initiative.
The GLM-5.3 model, built on the foundation of its predecessor GLM-5.2, boasts an impressive 743 billion parameters and aims to improve token efficiency significantly. The model is now available through subscriptions, with API access and downloadable weights to be rolled out following safety evaluations.
Performance and Efficiency Improvements
Z.ai emphasized that the development of GLM-5.3 focused on enhancing token efficiency instead of just performance. The model excels in generating results with fewer tokens, achieving a score of 34.5% on the internal Z.ai Code Bench—an improvement over GLM-5.2’s 23.4% but still trailing behind some proprietary models like Claude Fable 5, which scored 39.5%.
In coding capabilities, GLM-5.3 outperformed the Chinese model Kimi K3 in key benchmarks. It scored 28.3 in the Terminal Bench 3.0 test, while in the DeepSWE v1.1 benchmark, it earned 66.9, slightly below Kimi K3’s 67.5.
Notable Advancements in Cybersecurity
On the cybersecurity front, GLM-5.3 significantly improved performance, achieving an 84.5% score on CyberGym and identifying 2,436 vulnerabilities in various open-source projects, over a thousand of which were rated as medium-to-high severity. Z.ai noted the model has made considerable gains compared to GLM-5.2, especially in exploitation benchmarks.
The company described GLM-5.3 as taking "agentic coding to the next level" with its enhanced efficiency and performance.
Competitive Pricing and Future Releases
Z.ai also highlighted the model’s affordability, which undercuts prices of U.S. frontier models. The GLM Coding Plan utilizes a points system for billing, making it a cost-effective option compared to American alternatives. For example, GLM-5.2's pricing was significantly lower than that of models like GPT-5.3-Codex and others.
Despite being listed on the U.S. Entity List, Z.ai's GLM models show increasing popularity, with Chinese open-weight models often outperforming their U.S. counterparts in token utilization. The weights for GLM-5.3 are expected to be made publicly available within the next two weeks.
