Z.AI's GLM-5.3: Top Open-Weight Coding AI Model Released

Chinese AI lab Z.ai has released GLM-5.3, which it calls the "most capable open-weights model for coding." This 743-billion-parameter model is available through the GLM Coding Plan and ZCode. What's interesting is, Z.ai focused on token efficiency. GLM-5.3 uses fewer tokens per task compared to its predecessor, GLM-5.2. For example, on the Z.ai Code Bench, it scores 34.5% at Max effort using roughly 75,000 output tokens, while GLM-5.2 scored 23.4% with 96,000 tokens. While it outperforms its predecessor and some open-source peers, the blog notes it remains behind some closed models like Claude Fable 5 in certain benchmarks. However, in cybersecurity, GLM-5.3 shows a significant leap, leading CyberGym at 84.5% and more than doubling GLM-5.2 on exploitation benchmarks. It flagged 2,436 vulnerabilities across open-source projects. The bottom line is this new model offers improved performance and efficiency, especially in cybersecurity, which could impact how developers and security professionals work.