GLM-5.3: Z.ai Unveils New Model with Cybersecurity Focus

On August 14, 2026, Z.ai (formerly Zhipu AI, Beijing) officially introduced GLM-5.3, a new version of its flagship GLM-5 language model series, with the slogan 'Built to Code. Ready for Cyber Defense.' The model is built on the same 743-billion MoE architecture as GLM-5.2 but underwent additional post-training aimed at agentic programming and cybersecurity. Z.ai claims significant improvements in agentic coding and a new standard among open models for cyber defense, though specific benchmarks have not been published yet. GLM-5.3 is immediately available through the GLM Coding Plan subscription and the ZCode agentic IDE. API access and open weights will be released in stages after safety evaluations, a notable difference from GLM-5.2, whose weights were released under MIT license shortly after launch. A limited group of partners already offers services based on GLM-5.3 with safeguards. The release follows GLM-5.2's success, which achieved 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro, surpassing models like GPT-5.5. Independent tests by Semgrep and Irregular highlighted GLM-5.2's cyber capabilities, likely prompting Z.ai to emphasize cybersecurity in the GLM-5.3 announcement. The staged access is seen as a balance between openness and managing risks associated with enhanced offensive/defensive skills. As of now, exact benchmarks, context window, license terms, API pricing, and release dates for open access remain unknown.