Z.ai released GLM-5.3 on August 14, reusing the same 743-billion-parameter base model that powered GLM-5.2 rather than training a new one from scratch. Every reported capability gain comes purely from scaled-up post-training: more task environments, more environment variety, and longer reinforcement learning runs on top of the existing base. That is itself a notable engineering signal, since it suggests meaningful capability jumps no longer require the enormous cost of a fresh pretraining run, which matters for any team weighing how often to expect step-change releases from open-weight labs. The practical gains cluster in two areas that matter directly to software builders. On coding, long-horizon benchmarks moved the most: Terminal-Bench 3.0 jumped from 4.6 to 28.3 and DeepSWE v1.1 moved from 46.2 to 66.9, indicating the model got much better at multi-step, long-running agentic coding tasks like repository-scale refactors rather than single-shot code generation. On an internal Z.ai coding benchmark, the company reports roughly a 50% improvement over GLM-5.2. The second, more unusual result is in cybersecurity, which Z.ai says it did not specifically plan for: the model started forming coherent, multi-step exploitation plans as training scaled, not just isolated bug-spotting. On CyberGym, a white-box vulnerability discovery and validation benchmark, GLM-5.3 hit 84.5%, edging past Anthropic Mythos 5 at 83.8%. On ExploitBench, which requires full root-cause reasoning plus a working exploit, it more than doubled from 24.4% to 54.4%, though it still trails Mythos 5 at 78%. Because of that cybersecurity jump, Z.ai is delaying the public release of model weights for roughly two weeks while it completes safety hardening and evaluation, even though the model is already live through the Z.ai API, the GLM Coding Plan, and ZCode. For engineering teams, that combination is worth tracking on two fronts: GLM-5.3 is a legitimate, inexpensive option for long-horizon coding agents and CI failure triage, and its security-research capability is a preview of how open models are increasingly usable for both defensive vulnerability discovery and, with fewer guardrails than a closed frontier lab would apply, offensive tooling.