GLM-5.3 Open Weights Released

Z.ai released the GLM-5.3 open-weight model, which uses the same base model as GLM-5.2 but maximizes performance through post-training.

Improved Coding and Agent Performance

GLM-5.3 achieved over 50% performance improvement compared to GLM-5.2 in complex coding and long-horizon tasks. It recorded the highest performance among open-weight models on the internal benchmark Z.ai Code Bench, and achieved open-source SOTA (State-of-the-Art) on public benchmarks such as Terminal Bench 3.0 and Agents' Last Exam.

Rapid Advancement in Cybersecurity Capabilities

With the expansion of post-training scale, cybersecurity capabilities advanced faster than expected. It achieved state-of-the-art results on CyberGym, a vulnerability discovery benchmark, showing notable strengths at the top of the attack chain, particularly with performance improving by more than 2x compared to GLM-5.2 in the exploitation stage.

Deployment and Usage Guide

GLM-5.3 supports major inference frameworks such as SGLang, vLLM, Transformers, and Unsloth, and can be deployed on the Ascend NPU platform. During inference, the thinking budget can be adjusted via the reasoning_effort parameter, with a default value of max. In chat scenarios, clear_thinking=true must be explicitly passed.