What happened?
On August 14, 2026, Z.ai announced its new AI model, GLM-5.3. Instead of retraining the model, the company used the same 743-billion-parameter base model as GLM-5.2, attributing all performance gains to post-training scaling. Z.ai stated that during this process it applied more task environments, greater environment diversity, and longer training.
In coding tests, the biggest leap appeared in long-horizon tasks. Terminal-Bench 3.0 score rose from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and Agents' Last Exam (CLI) from 23.8 to 28.5. In the GDPval-AA v2 evaluation, which covers 44 professional categories, the model scored 1,769 points. In Z.ai's internal benchmark, Z.ai Code Bench, a 50% improvement over GLM-5.2 was reported.
An unexpected development in cybersecurity
Z.ai explained that the cybersecurity result was not planned. The company added vulnerability discovery data solely to improve isolated bug reasoning, but as training scaled, the model's capability grew exponentially. The model began generating consistent plans across end-to-end exploitation chains.
On CyberGym, which tests vulnerability discovery and verification from white-box source code, the score rose from 77.2% to 84.5%, surpassing Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). On ExploitBench, which requires root-cause analysis and working exploits, the score climbed from 24.4% to 54.4%; Mythos 5 remained ahead in this test with 78.0%.
What's known
- GLM-5.3 is available through the Z.ai API, GLM Coding Plan, and ZCode.
- Model weights have not yet been released; they are planned to be opened within about two weeks following safety evaluation and hardening.
- On ExploitGym, GLM-5.3 completed 105 tasks in two hours and 130 in six hours; GLM-5.2 completed only 29 and 39 tasks in the same timeframes.
- Mythos 5 remains ahead on ExploitGym with 181 tasks in two hours and 247 in six hours.
Why it matters
Z.ai's approach shows that significant capability gains can be achieved through post-training alone, without retraining the base model. This stands out as a method that could reduce model development costs. However, the unintended progress in cybersecurity also raises the need for policy and risk assessment within the security community.
The company notes that sectors such as developer tools, cloud infrastructure, application security, fintech, and e-commerce engineering could benefit from the model. Organizations with data locality requirements or vendor audit rules are advised to wait for the weights to be released.
Comparison Table
| Benchmark | GLM-5.2 | GLM-5.3 | Competing Model |
|---|---|---|---|
| Terminal-Bench 3.0 | 4.6 | 28.3 | — |
| DeepSWE v1.1 | 46.2 | 66.9 | — |
| CyberGym | 77.2% | 84.5% | Mythos 5: 83.8% |
| ExploitBench | 24.4% | 54.4% | Mythos 5: 78.0% |
What's next?
Z.ai plans to release the model weights for GLM-5.3 about two weeks after the announcement, once safety evaluation and hardening processes are complete. The company stated that all benchmark results are documented in the announcement along with the harness, context length, and sampling settings used.