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

Chinese AI developer Z.ai has released GLM-5.3, an open-source large language model that the company says set records on several common benchmarks. It builds on GLM-5.2, which Z.ai released in mid-July, and keeps the same mixture-of-experts setup with 753 billion parameters and a 1 million-token context window.

The big change is in the post-training. Z.ai says GLM-5.3 went through a more extensive refinement process, and that the result was a clear jump in coding performance. The model took the top score among open-source systems on Terminal Bench 3.0, a test for command-line scripting, and it did 50% better than GLM-5.2 on an internal benchmark for coding agents.

Security work is another focus. Z.ai says GLM-5.3 beat Claude Mythos 5 on CyberGym, a benchmark for finding code vulnerabilities, although it trailed Anthropic’s flagship model on two other cybersecurity tests. The company says the model has already uncovered more than 2,400 vulnerabilities across 269 software projects, with about half rated medium severity or higher. One of those bugs, Z.ai says, was in code written 40 years ago.

To get there, Z.ai built sandboxes meant to mimic developer workstations and had GLM-5.3 tackle complex coding tasks inside them. Some of those exercises ran for days, which is how the company says it trained the model for long-horizon work. Special AI agents generated the sandboxes, modeled them on real software projects, and created programming challenges for each one. A separate judge agent checked that the tasks were actually solvable before GLM-5.3 ever saw them.

Z.ai also automated parts of the training stack, including pipelines that generated reward signals to steer learning. The company says it built that stack on two open-source tools, slime and SAO. Slime helps move a model from training into production inference, while SAO is an asynchronous reinforcement learning implementation designed to speed up training runs. GLM-5.3 is available now through Z.ai’s GLM Coding Plan, and the company says the weights will be published on Hugging Face under an open-source license within two weeks.