Z.ai, the Beijing-based lab formerly known as Zhipu AI, released GLM-5.3 on August 14, the latest in a run of Chinese open-weight models aimed squarely at Anthropic and OpenAI’s coding lead.
Same base model, heavier post-training
Unlike most model upgrades, GLM-5.3 doesn’t come from a larger or freshly pretrained network. It sits on the same roughly 700-billion-parameter base model as June’s GLM-5.2, with Z.ai instead pouring resources into post-training — reinforcement learning and fine-tuning aimed specifically at coding and long-horizon agentic tasks. The company says the approach delivered a 50% improvement over GLM-5.2 on its internal coding-agent benchmark, with the largest gains on the hardest, longest-running tasks: its score on Terminal-Bench 3.0, which tests multi-step terminal and tool-use work, rose from 4.6 to 28.3.
A cybersecurity benchmark stands out
The most notable — and most double-edged — result is on CyberGym, a benchmark that scores a model’s ability to find and exploit software vulnerabilities. GLM-5.3 scored 84.5%, more than doubling GLM-5.2’s mark and edging out Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol on the same test, according to Z.ai’s own reporting cited by MarkTechPost and Silicon Republic. Stronger exploit-finding ability is useful for legitimate security research and automated patching, but it also raises the familiar dual-use concern that has followed other frontier coding models this year: a model that’s better at finding exploits is better for both defenders and attackers.
Availability and what’s still missing
GLM-5.3 is live now through Z.ai’s API and its GLM Coding Plan, rolled out automatically to existing coding-plan subscribers. The model’s weights — which would let developers download and run GLM-5.3 themselves, as they can with GLM-5.2 — are not yet public; Z.ai says it plans to release them within about two weeks, after further safety evaluation given the jump in exploit capability.
The release fits a pattern this year of Chinese labs — Z.ai, DeepSeek7.0 out of 10, Moonshot AI — competing on post-training efficiency and coding-agent performance rather than raw model size, betting that squeezing more capability out of an existing base model is cheaper and faster than a full retrain.