China’s AI labs sharpen open-model competition

China's artificial intelligence landscape is witnessing a surge as key players unveil advanced open-model systems. Z.ai unveils GLM-5.3, an AI coding model with robust cybersecurity capabilities, while DeepSeek introduces V4 Pro, an upgraded version of its large open-weight model. GLM-5.3 demonstrates a striking ability to identify software vulnerabilities, scoring 84.5% on CyberGym, a benchmark for code inspection and security weakness detection, compared to Anthropic's restricted-access Mythos 5 at 83.8%.

However, GLM-5.3 falls short in developing exploits from those vulnerabilities, with a score of 54.4% on ExploitBench, compared to Mythos 5's 78%. This discrepancy highlights the dual-use nature of advanced coding systems, where vulnerability discovery models can aid developers but also pose risks for attackers. Z.ai has delayed the full public release of GLM-5.3 for security assessments and will restrict sensitive cybersecurity functions to verified users under a trusted-access program.

DeepSeek's V4 Pro emphasizes autonomous agents, coding, tool use, and complex reasoning, boasting 1.6 trillion total parameters and a one-million-token context window. The model's pricing, at $1.32 per million input tokens and $3.96 per million output tokens, is significantly higher than its smaller, faster counterpart, V4 Flash.

As China's model market becomes increasingly crowded with capable systems from Z.ai, Moonshot AI, Alibaba, MiniMax, ByteDance, and others, DeepSeek is expanding its offerings beyond low-cost reasoning models towards agents, long-context applications, commercial APIs, and specialized infrastructure.