Zhipu AI launches GLM-5. 3 with enhanced coding and cyber capabilities

Zhipu AI has released GLM-5. 3, a new large language model that achieves significant performance gains in programming and cybersecurity through expanded post-training rather than architectural changes. Using the same 744-billion-parameter base as GLM-5. 2, the model reportedly shows a 50% improvement in internal coding benchmarks and has reached state-of-the-art levels on open-weight benchmarks such as Terminal Bench 3. 0 and Agents’ Last Exam.

The model has also demonstrated emergent cybersecurity capabilities. In CyberGym testing, GLM-5. 3 achieved an 84.5% success rate, surpassing Anthropic’s Mythos 5 and OpenAI’s GPT-5. 6 Sol. While it excels at identifying vulnerabilities, reports indicate it still trails competitors in complex exploit chain execution on the ExploitBench metric. During testing with security teams, the model helped identify 2,436 vulnerabilities across 269 projects, including one flaw dating back to 1981.

GLM-5. 3 is currently available via API and the GLM Coding Plan. Zhipu AI plans to release the full model weights to the public within approximately two weeks, following safety evaluations and hardening processes.