- Published on
Better at Complex Coding and Long-Horizon Tasks
- Authors
- Name
- DeepHorus
Z.ai has released GLM-5.3, a new model that promises major gains in complex coding and long-horizon tasks without retraining the underlying base model. It runs on the same 743B-parameter foundation as GLM-5.2, with every reported improvement coming from scaled post-training. The company expanded task environments and extended training, pushing the model further in reasoning and cybersecurity.
The release departs from the usual practice of retraining model weights. By adding more environment types and longer training to the post-training phase, Z.ai achieved significant benchmark jumps while keeping the base unchanged. This approach cuts compute costs and accelerates iteration.
Coding results show the largest gains on long-horizon benchmarks. Terminal-Bench 3.0 jumped from 4.6 to 28.3, while DeepSWE v1.1 improved from 46.2 to 66.9. On Z.ai Code Bench, an internal evaluation, GLM-5.3 scored 31.4% at 50,000 output tokens per task, a 50% improvement over GLM-5.2.
The cybersecurity result surprised the company. CyberGym, which tests white-box vulnerability discovery, rose from 77.2% to 84.5%, surpassing Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). ExploitBench more than doubled from 24.4% to 54.4%, though it trails Mythos 5 at 78.0%.
GLM-5.3 is live via the Z.ai API, GLM Coding Plan, and ZCode, but weights will only be published in about two weeks after safety evaluation and hardening. Startups and mid-market engineering orgs can adopt it now, while enterprises with data-residency rules should wait for the open weights. The company said the weights will be posted on GitHub and its security disclosure ledger.
The model targets developer tooling, cloud infrastructure, application security, and fintech/e-commerce engineering. Applications include repository-scale refactors, long-horizon CLI agents, CI failure triage, and white-box vulnerability discovery. Security vendors and MSSPs will likely feel the most impact, but they also face policy exposure.
Z.ai said the security boost was unplanned. It had added vulnerability-discovery data expecting better single-bug reasoning; instead, capability compounded as training scaled, leading to coherent exploit chains. All figures are vendor-reported, with harness and sampling settings documented.
With weights arriving soon, GLM-5.3 highlights a trend toward scaling post-training rather than base model size. Z.ai argues its private benchmarks reduce contamination risk. External validation in the coming weeks will determine whether these results hold.
