Zhipu releases GLM-5.3 with big coding boost, plans open source

Zhipu released its new open-source model GLM-5.3 today. It uses the same base as GLM-5.2 but gets a big boost from better post-training. The model is much stronger at coding, cybersecurity, and long tasks. Coding ability improved by 50%, and it tops open-source models on benchmarks like Terminal-Bench 3.0.

Scores jumped from 4.6 to 28.3 on Terminal-Bench 3.0, 46.2 to 66.9 on DeepSWE v1.1, and 23.8 to 28.5 on Agents' Last Exam. In Z.ai Code Bench, GLM-5.3 hit 31.4% accuracy, beating Claude Opus 4.8's 29.5%, while using fewer tokens. Now it's available in Zhipu's coding tool ZCode and efficiency tool AutoClaw. API is coming soon, and full weights will be open-sourced in two weeks after safety checks. Zhipu says GLM-5.3 is the top open-source model right now, close to Claude Fable 5 for coding and agent tasks.