China based AI-startup Zhipu AI (Z.ai), known for creating open-source General Language Model (GLM) family of LLMs, announced launch of its latest flagship model GLM-5.3, an open model post-trained on a 743B base, setting a new state-of-the-art (SOTA) benchmark among open source AI models.
According to Zhipu AI, GLM-5.3 AI model is better at complex coding and software development, long horizon tasks and cybersecurity compared to its previous version GLM-5.2.
Zhipu AI, considered one of six China’s AI Tiger companies, claimed the latest flagship model GLM-5.3 was deployed to several security teams in China for testing and it found 2,436 vulnerabilities across 269 projects, including 1,097 medium-to-high severity issues.
Introducing GLM-5.3: Built to Code. Ready for Cyber Defense.
– Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model
– A major leap in cybersecurity, setting a new standard among open modelsTech Blog: https://t.co/ekQkO83jCv pic.twitter.com/y3Y2AB0wxr
— Z.ai (@Zai_org) August 14, 2026
“The findings span system kernels, operating systems, browser engines, open-source infrastructure, web applications, and network protocols. Many had remained unnoticed for years or even decades, with the oldest dating back roughly 40 years,” claimed Zhipu AI in a statement on Friday.
The company says GLM-5.3 uses the exact base model as GLM-5.2: a ~743–744 billion parameter Mixture-of-Experts (MoE) architecture with roughly 40 billion parameters activated per token. It supports a 1 million token context window and up to 128K output tokens. It is text-only.
“GLM-5.3 is the most capable open-weights model for coding, with a 50% improvement over GLM-5.2 on our in-house Z.ai Code Bench. It also achieves open-source SOTA on public benchmarks including Terminal Bench 3.0 and Agents’ Last Exam,” claimed Zhipu AI in a blog post on Friday.
The company also claimed that its maintaining a public security disclosure ledger regarding the findings made by GLM-5.3.
“The ledger is continuously updated as new vulnerabilities are reviewed and disclosed, distinguishing issues that have already been made public from those that are still under disclosure. For disclosed issues, it records information including the affected project, severity, CVE where available, and how long the vulnerability had remained in the codebase,” said the company.
The AI company has claimed that GLM-5.3 outperforms or matches leading closed models such as Anthropic’s Mythos 5 (also referred to as Fable 5 in some evaluations) and OpenAI’s GPT-5.6 Sol on key cybersecurity benchmarks.
The company also positions the model as competitive or superior in certain coding and agentic tasks relative to models like Kimi K3, DeepSeek-V4, and some Claude Opus variants, particularly in efficiency and long-horizon performance, while describing it as the strongest open-weights coding model overall.
These performance figures come exclusively from Z.ai’s own internal testing and vendor-reported results published in its technical blog and announcements. No independent third-party assessments or external verifications of the full benchmark suite has been carried out yet.
Also Read: Zhipu AI Market Cap Tops HK$1 Trillion as Shares Soar 42%
