Z.ai's GLM-5.3 boosts coding and cybersecurity performance without adding parameters

Z.ai released GLM-5.3 in August 2026, achieving a 50% improvement in programming capabilities over its predecessor GLM-5.2 while keeping the same 743 billion parameters and base architecture unchanged. The gains came entirely from post-training refinements, including better learning methods, improved data quality, and expanded reinforcement learning at industrial scale. GLM-5.3 ranked first globally on the CyberGym vulnerability detection benchmark with an 84.5% score and topped several open-source coding benchmarks, while also uncovering 2,436 vulnerabilities across 269 real-world projects, including a DNS protocol bug dormant since 1983. Z.ai plans to open-source the model weights within two weeks, alongside a controlled-access program and a community-driven security initiative. The results challenge the prevailing assumption that larger models require more parameters, suggesting that post-training innovation can deliver comparable or greater gains at lower inference costs.