China's Z.AI Ships GLM-5.3, Calling It the Top Open-Weight Coding Model
Z.ai released GLM-5.3, a 743B-parameter coding model it says is the strongest open-weights coder coming soon. It is currently available through the GLM Coding Plan and ZCode, with API access and downloadable weights to follow after safety review. The main improvement is post-training scaling: better agentic coding with fewer output tokens per task. Z.ai says GLM-5.3 scores 34.5% on its Z.ai Code Bench while using about 75k output tokens, versus GLM-5.2’s 23.4% at 96k. It outperforms several open peers, including Kimi K3 on key coding benchmarks, but still trails top closed U.S. models on major boards. It also shows a large cybersecurity jump, leading CyberGym at 84.5% and finding thousands of vulnerabilities in open-source projects. Pricing is much lower than U.S. frontier models, and public weights are expected in about two weeks.
