Zhipu releases GLM-5.3 with 50% coding boost from post-training alone, tops open-source Terminal-Bench
Zhipu Z.ai released GLM-5.3 on August 15. Using the same 743B MoE base, pure post-training drove a 50% coding gain over 5.2, open-source SOTA on Terminal-Bench 3.0, and surpassed Claude Mythos 5 on CyberGym.
Z.ai (Zhipu) released GLM-5.3 open-source large language model in the early hours of August 15. The version uses the same 743B MoE base as GLM-5.2, and pure post-training drove a roughly 50% improvement on coding agent benchmarks. On Terminal-Bench 3.0, GLM-5.3 took the top spot among all open-source models.
On the cybersecurity side, GLM-5.3 surpassed Anthropic Claude Mythos 5 on the CyberGym benchmark. Through the Project Glasswing effort, GLM-5.3 discovered 2,436 unpatched open-source vulnerabilities, of which 1,097 are rated critical or high severity, spanning a 26-year historical range. This is the first time an AI model has delayed open-source release pending a cybersecurity review—Zhipu announced a two-week delay to complete the review and launched the cvd.z.ai vulnerability disclosure platform.
GLM-5.3 is a landmark case for the "post-training scaling" path, showing significant capability gains from high-quality post-training data even without changing the base model. In the same week, Alibaba Qwen open-sourced Qwen3.8-27B, DeepSeek Harness launched, and ByteDance established an AI Data & Safety first-level department.