Z.ai has introduced GLM-5.3, an open-source large language model that has made significant advancements in coding and cybersecurity. GLM-5.3 is an evolution of GLM-5.2, featuring a mixture of experts architecture with 753 billion parameters and a context window of 1 million tokens. The model underwent an extensive post-training process, resulting in superior performance across various benchmarks.
One standout achievement of GLM-5.3 is its success on Terminal Bench 3.0, outperforming all other open-source AI models in command line scripting capabilities. Additionally, it excelled in cybersecurity research, surpassing Claude Mythos 5 on CyberGym in identifying code vulnerabilities. However, it lagged behind Anthropic’s flagship LLM in other cybersecurity evaluations.
Z.ai reports that GLM-5.3 has uncovered over 2,400 vulnerabilities in 269 software projects, with many flaws rated as medium to high severity. Notable is the discovery of a vulnerability in code authored four decades ago, showcasing the model’s comprehensive analysis capabilities.
The training process for GLM-5.3 involved intricate exercises within simulated developer workstations. Z.ai deployed specialized AI agents to create realistic environments and custom programming challenges. By automating aspects of the training workflow and utilizing reward signals for guidance, the company enhanced the model’s coding proficiency and problem-solving abilities.
Z.ai leveraged the slime and SAO open-source technologies to streamline the model’s transition from training to production. The incorporation of asynchronous reinforcement learning through SAO accelerated training runs. GLM-5.3 is currently accessible through Z.ai’s GLM Coding Plan subscription service, with plans to release its weights on Hugging Face under an open-source license soon.
