GLM-5.3

Zhipu AI released GLM-5.3 on 14 August 2026 under the banner "Built to Code. Ready for Cyber Defense." Unusually, the base model is unchanged from GLM-5.2 — Zhipu attributes the whole gain to heavy post-training scaling, and reports a 50% improvement in coding capability on internal evaluations. It ranks first among open-source models on Terminal Bench 3.0 and Agents' Last Exam, and posts 84.5% on the CyberGym vulnerability-reasoning evaluation. The model offers a 1M-token context with 128K maximum output, and Zhipu open-sourced the weights roughly two weeks after the API release.