GLM 5.3 is a frontier coding-focused model from Z.ai that shares its base architecture with the previous GLM 5.2 release and draws every reported improvement from scaled post-training rather than from a new pretraining run. The post-training stack rests on three research pillars developed for the prior version: IndexShare for efficient long-context processing, SAO for reinforcement learning on long-horizon tasks, and the slime framework for large-scale asynchronous training, all exercised over an expanding library of task environments and compute. Rather than rebuilding the foundation, the team concentrated on longer and more diverse agentic environments to push the same base toward harder software and security work.
The result is a model positioned for complex software engineering and emerging cybersecurity use cases. On the Z.ai in-house Code Bench, GLM 5.3 posts a roughly fifty percent lift over GLM 5.2, and it reaches open-source state-of-the-art results on public benchmarks such as Terminal Bench 3.0 and Agents' Last Exam, with additional strong showings on DeepSWE v1.1, NL2Repo, ProgramBench, FrontierSWE, SWE-Marathon, and PostTrainBench. Cyber capability scaled faster than expected during post-training, putting GLM 5.3 at the top of CyberGym for vulnerability discovery and more than doubling GLM 5.2 on exploitation benchmarks, which makes it a practical fit for teams that want a single open-weights model for long-horizon coding agents, code review, and security-oriented research workflows.
