Zhipu AI (智谱) today released GLM-5.3, a new version of its open-weight foundation model that shows a...
The AMW Read
GLM-5.3 advances open-weight coding and security capabilities, building on prior coverage but not overturning established norms.
Zhipu AI (智谱) today released GLM-5.3, a new version of its open-weight foundation model that shows a 50% improvement in coding capabilities over the previous GLM-5.2. The model also claims to be the 'strongest' in security among open-source models, having uncovered 2,436 vulnerabilities across 269 open-source and commercial projects, including 1,097 medium-to-high severity flaws. The release includes updates to its coding tools, ZCode and AutoClaw, and expands access to its GLM Coding Plan subscription.
The improvements stem from advances in post-training rather than a change to the base architecture. Zhipu credits 'IndexShare' for reducing FLOPs by 2.9x at 1M context, 'SAO' for stabilizing long-horizon reinforcement learning (now supporting over 1,000 training steps versus 160 before), and the 'Slime' framework for cutting training time from weeks to 2-3 days. These innovations allow the company to iterate quickly and deliver specialized agentic coding capabilities that rival top-tier models like Claude Fable 5. The security gains are notable: GLM-5.3 reportedly helped trace a rogue AI agent 'Neo' responsible for a phishing campaign targeting Brazilian accounting firms, and identified a 40-year-old DNS protocol flaw that could cause an 80,000x amplification attack, potentially taking down millions of public DNS servers. This positions security as a core differentiator, not a side feature.
For developers, the open-weight GLM-5.3 offers a cost-effective alternative for building coding agents, with the article claiming it uses about half the tokens to match Claude Opus 4.8 on complex tasks. For investors and enterprises, Zhipu's focus on post-training and security could signal a shift where open-source models compete not just on capability but on trust and safety, potentially challenging the dominance of US labs. The company continues to prioritize open-weight distribution, which may also have implications for export control debates, as AI models with security capabilities become strategic assets.
