Zhipu AI Launches GLM-5.3 to Challenge Frontier SOTA Models in Coding and Cybersecurity

Artificial intelligence company Zhipu AI (Z.ai) has officially released GLM-5.3, introducing substantial upgrades to its flagship model series. The release keeps the base architecture of GLM-5.2 while delivering significant performance gains through post-training across software engineering, terminal execution, and digital security.

Faster, More Capable Coding

GLM-5.3 shows marked progress in solving complex, multi-step engineering projects from start to finish. On Zhipu’s internal coding evaluations, the model performed ∼50 better than GLM-5.2, resolving software engineering challenges with higher accuracy while using fewer computational resources.

The model also show strong results across standard industry benchmarks:

• DeepSWE v1.1: Jumped from 46.2 to 66.9 in solving real-world software bugs.

• FrontierSWE: Reached 78.1, up from 67.5.

• Terminal Bench 3.0: Climbed from 4.6 to 28.3, demonstrating a much stronger ability to navigate command-line interfaces and interactive developer tools.

Advanced Cybersecurity Capabilities

Three weeks ago, Hugging Face deployed its own GLM model to perform forensic log analysis, dissect complex attack payloads, and audit vulnerabilities during major security incidents.

GLM-5.3 further improved these capabilities. The new model was trained in interactive simulation environments to discover, analyze, and verify complex vulnerabilities end-to-end:

• CyberGym: Scored 84.5, placing it among the highest-ranked AI models in security evaluations.

• ExploitBench: Surged to 54.4, more than doubling the score of its previous generation (24.4).

• ExploitGym: Solved 130 complex security challenges in 6 hours, compared to just 39 on GLM-5.2.

Built-in Reasoning and Training Upgrades

GLM-5.3 now works through problems step-by-step by default. This deliberate reasoning process significantly reduces careless mistakes on complex assignments. Users can easily control how much time the model spends analyzing a problem, switching between fast, lightweight answers for simple queries or deep, multi-step reasoning for demanding tasks.

These gains were made possible by major improvements to Zhipu’s AI training pipeline. By speeding up its learning systems by more than 2.3×, the team enabled the model to practice solving much longer and more realistic multi-stage scenarios during training.

Availability

GLM-5.3 is available starting today for developers and businesses through Z.ai's API and the ZCode coding platform. The open model weights of GLM-5.3 is expected to be publicly available in two weeks.