Chinese firm Zhipu AI claims its new GLM-5.3 open-weight model narrowly outperforms Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol on the CyberGym software security benchmark, following extensive internal evaluations. GLM-5.3 builds on a 743‑billion‑parameter mixture‑of‑experts base and reportedly detected 1,097 serious bugs across real-world projects, including Linux kernel, WinRAR, Redis, and FFmpeg, with some vulnerabilities allegedly overlooked for decades. Weights will be released on Hugging Face around August 28 after additional safety review, while access remains limited to paying subscribers. However, all reported benchmark scores come from Zhipu AI’s own tests rather than independent assessments.
This update represents a notable development in the Crypto sector. Organizations and founders tracking this space should evaluate potential strategic and technical implications on their operations.