Zhipu AI Lanza GLM-5.3: Lidera en Terminal Bench 3.0, CyberGym (84.5%) y AutomationBench

Por Dr. Erik Vandeberg • Publicado el 2026-09-11 • 5 min read • Fuente: Zhipu AI / Z.ai

GLM-5.3 logra el nuevo estado del arte open-source en Terminal Bench 3.0 y CyberGym. Desarrollado con pesos abiertos para entornos de ingeniería y agentes de ciberseguridad.

Zhipu AI (Z.ai) has released **GLM-5.3** and **GLM-5.3-Flash**, demonstrating emergent capabilities in automated software engineering and cybersecurity.

Unprecedented Cybersecurity & Vulnerability Discovery Scaling post-training on high-fidelity vulnerability simulations unlocked remarkable cyber capabilities in GLM-5.3: - 84.5% on CyberGym: The highest score recorded across all open and closed frontier models. - ExploitGym (6h): 130 successful vulnerability chains, more than doubling GLM-5.2's score of 39. - AutomationBench (v1.0.6): 48.2%, outperforming both Claude Opus 4.8 and GPT-5.6 Sol in real-world system automation.

Terminal Bench 3.0 Leadership On the notoriously difficult Terminal Bench 3.0—which evaluates multi-step bash debugging, environment recovery, and tool compilation—GLM-5.3 scored 28.3%, surpassing Kimi K3 (17.4%) and GLM-5.2 (4.6%).

Fuente y Verificación

Este informe técnico fue contrastado contra la documentación primaria publicada por Zhipu AI / Z.ai.

Leer Anuncio Original en Zhipu AI / Z.ai →

Zhipu AI Lanza GLM-5.3: Lidera en Terminal Bench 3.0, CyberGym (84.5%) y AutomationBench

Terminal Bench 3.0 Leadership On the notoriously difficult Terminal Bench 3.0—which evaluates multi-step bash debugging, environment recovery, and tool compilation—GLM-5.3 scored **28.3%**, surpassing Kimi K3 (17.4%) and GLM-5.2 (4.6%).

Fuente y Verificación

Terminal Bench 3.0 Leadership On the notoriously difficult Terminal Bench 3.0—which evaluates multi-step bash debugging, environment recovery, and tool compilation—GLM-5.3 scored 28.3%, surpassing Kimi K3 (17.4%) and GLM-5.2 (4.6%).