Z.ai has released GLM-5.3, an iteration of its large language model that leverages the existing 743B base model from GLM-5.2. All reported performance enhancements are attributed to scaled post-training, incorporating a broader array of task environments and longer training durations, rather than base model retraining. Partially, the model is currently accessible via the Z.ai API, the GLM Coding Plan, and ZCode. Public weights are anticipated in approximately two weeks, following safety evaluations and hardening.
Performance gains are particularly evident in coding and long-horizon tasks, alongside improvements in cybersecurity. Key benchmark results include:
- Coding: Terminal-Bench 3.0 improved from 4.6 to 28.3; DeepSWE v1.1 from 46.2 to 66.9; Agents' Last Exam (CLI) from 23.8 to 28.5. On GDPval-AA v2, which spans 44 occupations, GLM-5.3 scored 1,769. Z.ai's internal Code Bench reported a 50% improvement over GLM-5.2, with GLM-5.3 achieving 31.4% (~50,000 output tokens) against Claude Opus 4.8's 29.5% (120,000 tokens).
- Cybersecurity: CyberGym, which tests discovery and validation from white-box source, scores increased from 77.2% to 84.5%, surpassing Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). ExploitBench, which requires root-cause reasoning and a working exploit, saw a jump from 24.4% to 54.4%. On ExploitGym, GLM-5.3 completed 105 tasks in two hours and 130 in six hours, compared to GLM-5.2's 29 tasks in two hours and 39 in six hours.
These improvements, particularly in cybersecurity, were an unplanned outcome of scaled post-training, leading to the model forming coherent plans across complete exploitation chains rather than just single-bug reasoning. While GLM-5.3 demonstrates competitive results, especially in cybersecurity, it still trails GPT-5.6 Sol and Fable 5 on several harder public coding evaluations. The consistent pattern indicates larger gains over GLM-5.2 on benchmarks deeper into the exploitation chain.