China's AI industry has grown fast over the past few years. Companies like Z.AI, Moonshot AI, and DeepSeek are now close behind US leaders such as OpenAI and Anthropic in model performance.
Many of the founders behind these companies studied together at Tsinghua University in Beijing. Tang Jie, a professor there, co-founded Z.AI. His former student, Yang Zhilin, now leads Moonshot AI.
Industry leaders in both the US and China say the gap between the best American and Chinese models is now just a few months. This month, Z.AI released its GLM-5.3 model and said it matched Anthropic's Mythos 5 model in cybersecurity testing.
How Chinese Firms Built Their Models
Chinese developers used a mix of methods to catch up. These include hiring back Chinese researchers who trained in the US, following published research closely, and using open-source tools that are free to download and edit.
Some companies also used a process called distillation. This involves a new AI model learning from an older one by asking it many questions and studying the answers.
Anthropic has said Z.AI and Moonshot broke its usage policies through large-scale distillation. Neither company has responded publicly to the claim. People in the industry say distillation is common worldwide, though some Chinese firms may use it more than others.
DeepSeek built a technique called multihead latent attention. This method cuts down how much memory a model uses during a chat, which lowers computing costs.
DeepSeek also adopted a model design called mixture of experts. This setup sends a task to a specialized part of the system instead of using the whole model at once. It helped Chinese firms boost performance without needing as many advanced chips.
These methods were behind DeepSeek's release of a low cost model in January 2025, an event that caused a brief drop in US tech shares.
Government Support And Funding Gaps
China's government has backed AI research through state funds and university programs for decades. Nearly half of all equity investment in China this year went into AI, much of it from government backed funds.
Still, Chinese firms have less money to work with than their US rivals. Z.AI raised about $1.25 billion by mid-2025. Anthropic's revenue alone reached $65 billion in July, according to people briefed on the company's finances.
Chip access remains a hurdle too. Researchers at firms like Alibaba and Z.AI say they often get about one-fifth of the advanced chips that US labs receive.
In January, Z.AI became the first Chinese AI model startup to list its shares on the Hong Kong stock exchange. By July, its annual recurring revenue had reached $1 billion, roughly twice that of DeepSeek.
Tang is now training Z.AI's next model. He has told staff he wants it to handle complex research tasks that take weeks to complete.
