New Chinese AI Model Rivals Anthropic on Cybersecurity

The Latest:

Chinese AI startup Z.AI released its GLM-5.3 model this month, saying it matches Anthropic’s Mythos 5 in cybersecurity capability, as Chinese labs draw on DeepSeek-pioneered techniques to close the capability gap despite restricted chip access, the Wall Street Journal reported.

Details:

• Techniques: DeepSeek built multihead latent attention (MLA) to cut memory use and a mixture-of-experts (MoE) design to route tasks to specialized models, per WSJ.
• Chip gap: Researchers at labs like Alibaba and Z.AI get about a fifth of the advanced chips available to OpenAI and Google peers, they said.
• Timeline gap: Industry leaders in both countries say China’s best AI model is now just months behind the top US model, versus years previously.
• Distillation dispute: Anthropic accused Z.AI and Moonshot of extensively “distilling” its models by mining their answers to train rivals; neither company has commented.
• Chips, not talent: DeepSeek founder Liang Wenfeng said chip access, not people, is the biggest gap between Chinese and US AI firms.

By the Numbers:

Z.AI raised roughly $1.25 billion by mid-2025, far less than OpenAI and Anthropic have raised. Some Chinese firms are building models with 5 trillion to 10 trillion parameters, near Anthropic’s latest models, estimated above 5 trillion, per WSJ.

Between the Lines:

Tang Jie and his former student Yang Zhilin’s rise shows China’s AI leap wasn’t sudden — it grew from a research network built at Tsinghua University over more than a decade, the Journal reported.

What to Watch:

Tang is training Z.AI’s next flagship model for complex, weeks-long research tasks, people familiar with the plans said, while Anthropic’s distillation accusations against Z.AI and Moonshot remain unresolved.