GLM-5.3 pairs stronger coding with a sharp jump in cyber capability — and delays its open weights for safety review

On August 14, 2026, Z.ai announced GLM-5.3, a post-trained successor to GLM-5.2 that uses the same base model but substantially more reinforcement-learning work on long-horizon coding and security environments. Z.ai reports large gains on coding and exploitation benchmarks and says the model’s cyber capability developed faster than expected. GLM-5.3 is available through Z.ai’s hosted coding products, but the company says the model weights will not be published until roughly two weeks after launch while safety evaluation and hardening are completed.

The release is a useful signal for builders because it ties three trends together: open-weight models are becoming credible coding-agent backends, post-training can sharply change capability without a new base model, and cybersecurity performance can improve quickly enough that release mechanics themselves become part of the product decision. Teams evaluating self-hosted coding agents should not treat an announced open model as equivalent to immediately downloadable weights, and teams giving coding agents network or execution access should expect security capability to rise alongside general coding ability.

Z.ai says GLM-5.3 uses the same base model as GLM-5.2 and attributes the improvement to scaling post-training across more environments, tasks and reinforcement-learning compute. In the company’s published results, GLM-5.3 improves sharply on several coding and agent benchmarks, including Terminal Bench 3.0 and DeepSWE v1.1. Those are vendor-run results and should be treated as directional until independently reproduced under comparable harnesses and budgets.

Z.ai says vulnerability-discovery training produced larger gains as tasks moved further up the exploitation chain. Its published scores show especially large increases on exploitation-oriented benchmarks compared with GLM-5.2. The company also says deployments with security teams identified thousands of suspected vulnerabilities across open-source projects, although those findings are primarily first-party evidence and many remain under disclosure.

Unlike a conventional open-weight launch where model artifacts arrive with the announcement, Z.ai says GLM-5.3’s weights will be released about two weeks later after additional safety evaluation and hardening. Builders can use the model now through Z.ai’s hosted coding products, but local deployment and independent auditing must wait for the artifacts.

Z.ai’s GLM Coding Plan now uses a points-based quota that accounts separately for input, cached input and output tokens. Calls outside weekday peak hours consume half the normal points. That makes workload timing and cache behavior part of the effective economics for hosted coding-agent use, while the eventual open-weight release will introduce a separate self-hosting trade-off.