Z.ai Releases GLM-5.3 MoE Model with Scaled Post-Training and Cybersecurity Upgrades
Z.ai (Zhipu AI) officially released GLM-5.3 on August 14, 2026. The new release is a post-training refresh of the company's flagship 743-billion-parameter Mixture-of-Experts (MoE) architecture. All gains are attributed to scaled-up reinforcement learning and training on complex agentic environments, rather than changes to the underlying base model parameters or context window length.
According to vendor-reported benchmarks, the model achieves substantial performance improvements over its predecessor, GLM-5.2. On Terminal-Bench 3.0, which tests multi-step command line execution, GLM-5.3 scored 28.3% compared to just 4.6% for GLM-5.2. In cybersecurity tasks, its CyberGym benchmark score rose to 84.5% from 77.2%, showcasing a significant leap in safe scripting and exploit analysis capabilities.
Reasoning Controls and Licensing
GLM-5.3 introduces a standard reasoning_effort control parameter (low, high, max), with "max" set as the default option. In a departure from earlier model behaviors, Z.ai has locked the "thinking" mode on; calls attempting to disable internal reasoning trajectories will return api errors.
The model is text-only and is currently accessible to developers via the GLM Coding Plan and ZCode interface. Z.ai plans to release the open-weight model checkpoints under their custom license in late August 2026 following safety review procedures.
