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

Chinese AI firm Z.ai unveiled GLM-5.3, an open-source large language model that shattered benchmarks. GLM-5.3 is an iteration of the mid-July-released GLM-5.2, boasting 753 billion parameters and a context window of 1 million tokens. The model underwent extensive training optimizations that yielded impressive performance gains. GLM-5.3 outperformed all other open-source AI models on Terminal Bench 3.0, excelling in command line scripting tasks.

The model demonstrated a 50% improvement over its predecessor on Z.ai's coding agent evaluation benchmark. Notably, GLM-5.3 excels in cybersecurity research, outperforming Claude Mythos 5 on CyberGym, a benchmark that assesses a model's ability to detect code vulnerabilities. The model has discovered over 2,400 vulnerabilities in 269 software projects, with half of the flaws rated as medium or higher severity.

One of the vulnerabilities was found in code written 40 years ago. Z.ai refined the model's coding abilities through a post-training process that simulated developer workstations in sandboxes. The company generated these sandboxes using AI agents that modeled real-world software projects and created customized programming exercises for them.

A separate judge agent ensured the tasks were solvable before presenting them to GLM-5.3. Z.ai also streamlined the training workflow by automating certain aspects, such as generating reward signals to guide the LLM learning process. The company utilized open-source technologies slime and SAO to facilitate the model's transition from training to production.

GLM-5.3 will be available through Z.ai's GLM Coding Plan subscription service, with the model's weights released on Hugging Face under an open-source license within two weeks.