GLM-5.3 Benchmarks Show Major Leaps in Coding and Security

If you have been waiting for an open-weights model you can actually run locally, GLM-5.3 from Z.ai is starting to look like the one. The benchmark numbers just dropped, and they show huge jumps in coding and offensive security.

Those gains are concrete, and the weights land in two weeks.

What Happened

Z.ai shared early benchmark results for GLM-5.3, an upcoming open-weights model. The numbers are a clear step up from earlier versions:

• Terminal Bench 3.0: jumped from 4.6 to 28.3
• DeepSWE v1.1: rose from 46.2 to 66.9
• Internal Code Bench: about 50% higher than GLM-5.2
• AutomationBench: 48.2, topping the chart
• GDPVal-AA v2: 1769, also at the top

The security results are the standout. CyberGym scored 84.5, beating GPT-5.6 Sol's 83.6. ExploitBench hit 54.4, a 2x improvement. In real-world testing, the model found 2,436 vulnerabilities across 269 active projects.

Weights are scheduled to release in two weeks.

Why It Matters

This is a strong signal that scaling post-training matters as much as the architecture itself. Once the weights drop, you will be able to run a model locally that competes with closed models on coding and security work.