Z.ai debuts GLM-5.3 with long-horizon coding, cybersecurity upgrades

Chinese artificial intelligence developer Z.ai Co. today debuted GLM-5.3, an open-source large language model that set records across several popular benchmarks.

The LLM is based on an algorithm called GLM-5.2 that the company released in mid-July. The latter model features a mixture of experts architecture with 753 billion parameters and a context window of 1 million tokens. GLM-5.3 has an identical design, but went through a more extensive post-training process.

Z.ai says its training optimizations delivered significant performance improvements. GLM-5.3 achieved the highest score of any open-source AI model on Terminal Bench 3.0, which measures LLMs’ command line scripting capabilities. It performed 50% better than GLM-5.2 on an internal Z.ai benchmark for evaluating coding agents.

Notably, the model is also highly adept at cybersecurity research. It outperformed Claude Mythos 5 on CyberGym, a benchmark that evaluates LLMs’ ability to find code vulnerabilities. GLM-5.3 fell behind Anthropic’s flagship LLM on two other cybersecurity benchmarks. 

Z.ai says that the model has so far found more than 2,400 vulnerabilities in 269 software projects. About half of the flaws have a severity rating of medium or higher. According to the company, one of the vulnerabilities that GLM-5.3 found is in a piece of code authored 40 years ago.

The post-training process through which Z.ai refined the model’s coding capabilities involved sandboxes designed to mimic developer workstations. The company installed GLM-5.3 in the sandboxes and instructed it to complete complex coding tasks. Some of the exercises took days to complete, which improved the model’s ability to tackle long-horizon tasks.

Z.ai generated the sandboxes using specialized AI agents. According to the company, the agents modeled the environments they generated on real-world software projects. They then created programming exercises customized to each sandbox. A separate “judge agent” verified that the challenges  can be solved before they were given to GLM-5.3.

According to Z.ai, its engineers also automated certain other aspects of the training workflow. The company built pipelines capable of generating a reward signal, a piece of data that guides the LLM learning process. It provides feedback that helps the model being trained identify ways of refining its output.

Z.ai built its training stack on two open-source technologies called slime and SAO. The former tool makes it easier to move an LLM from its training environment to production inference infrastructure. SAO, in turn, is an implementation of an AI method called asynchronous reinforcement learning that speeds up training runs.

GLM-5.3 is currently available through Z.ai’s GLM Coding Plan subscription service. The company plans to release the model’s weights on Hugging Face under an open-source license within two weeks.

Image: Unsplash