Chinese artificial intelligence developer Z.ai Co. this morning officially launched the new version of GLM-5.3, an open-source large-scale language model that has already achieved impressive results in both programming and cybersecurity research. The model is based on the previous GLM-5.2 algorithm, released in mid-June, but has undergone extensive post-training in recent weeks, allowing for a significant performance boost.
Architecture and parameter scale
GLM-5.3 retains the same „mixture of experts“ architecture as GLM-5.2, but this time the model has 753 billion parameters and a huge context window of up to 1 million tokens. This scale allows the model to maintain long-term context, which is especially important for complex coding tasks that can last several days.
Long-term encoding improvements
In the post-training phase, Z.ai created sandboxes specifically designed to simulate programmers„ workstations. In these environments, the model was challenged to complete complex programming tasks, with some exercises lasting several days. These “long-term„ tasks helped GLM-5.3 learn to consistently plan and execute large code fragments, rather than just short fragments. In addition, the sandboxes were generated by AI agents that modeled real-world program structures and created customized programming tasks. Each task was checked by a “judge agent” before being submitted to the model for processing, ensuring that the task was feasible.
Cybersecurity research
GLM-5.3 not only excels in coding, but also demonstrates strong results in cybersecurity research. The model achieved the highest score among open-source AI models in Terminal Bench 3.0, a test that measures LLM’s ability to write command-line scripts. In the area of code security, GLM-5.3 outperformed Claude Mythos 5 in the CyberGym benchmark, which measures a model’s ability to detect vulnerabilities in applications. While the model fell slightly behind Anthropic’s latest LLM in some other cybersecurity benchmarks, it still proved capable of finding a large number of vulnerabilities.
Z.ai says that to date, GLM-5.3 has detected more than 2,400 vulnerabilities in 269 software projects, about half of which have a severity rating of medium or higher. One of the vulnerable code fragments found was written 40 years ago, demonstrating the model’s ability to analyze old and rare code bases.
Training infrastructure and accessibility
The model training process was facilitated by two open source tools, slime and SAO. Slime simplifies the migration of LLM from the training environment to the production inference infrastructure, while SAO, an asynchronous reinforcement learning method, speeds up training cycles. In addition, Z.ai created an automated data stream that generates a reward signal that acts as feedback to the model training, helping to continuously improve the output.
GLM-5.3 is currently available through Z.ai’s GLM Coding Plan subscription service. The company plans to release model weights on the Hugging Face platform in two weeks, enabling the community to use and further develop this large-scale language model.
Conclusions
Z.ai GLM-5.3 is a major step forward in both AI programming and cybersecurity. The huge number of parameters, huge context window, and comprehensive post-training process give the model the ability to solve complex, long-term tasks, while simultaneously detecting and analyzing application vulnerabilities. The open-source license and rapid release of weights further encourage collaboration and innovation in the technology community.
—
CTA: If you want to learn how GLM‑5.3 can help your organization strengthen cybersecurity or automate programming processes, contact Krikis IT specialists.
If you want to learn how GLM‑5.3 can help your organization strengthen cybersecurity or automate programming processes, contact Krikis IT specialists.
