(Yicai) Sept. 1 -- Over the last two months, there have been a series of frontier model releases in China, from Moonshot AI’s Kimi K3 to Z.AI’s GLM-5.3-Flash, and from Alibaba Group Holding’s Qwen 3.8-Max to DeepSeek’s DeepSeek-V4-Pro.

But when compared to the updated technical benchmark rankings, a handful of recent business announcements are even more revealing about the direction of Chinese AI.

Harvey, a US legal artificial intelligence firm backed by OpenAI, Sequoia Capital, and Andreessen Horowitz, released its first in-house model, which was post-trained on Kimi K3. A day later, Thomson Reuters unveiled its own proprietary model, built at a reported cost of USD40 million on Qwen. Last week, Alibaba opened an international edition of its AI agent platform QwenWork, making it available to users outside China.

These announcements are not particularly novel. Cursor's coding model was built on Kimi K2.5. Airbnb leans on Qwen, and Perplexity on DeepSeek. In June, Lindy switched from Anthropic to DeepSeek. Some of the most sophisticated software companies in the United States are constructing their products on Chinese foundations.

This is happening because AI is open-weight, low-cost, and application-focused in China. Firms monetize through adjacent services, but broadly speaking, the models can be downloaded and run on local servers. Efficiency gains, coming in part from US export restrictions, mean that Chinese models can achieve high performance at a lower compute cost. Meanwhile, the US’s top AI models have metered access, driving up massive budgets for token spending.

The bottom line is that Chinese firms are offering top-shelf AI at a fraction of the cost, and open-weight allows for customization.

Part of this divergence is that US labs have stated they are pursuing artificial general intelligence, while Chinese firms are more focused on creating models that can be diffused across the economy via customized applications. For example, Alibaba’s Chairman Joseph Tsai said open source is important because it lets a wide range of companies and entrepreneurs innovate in AI, rather than letting a few companies own all the windfall.

My own firm, WPIC Marketing + Technologies, is one such example. We help global consumer brands sell in China, Japan, South Korea, and Southeast Asia, running their online stores, warehouses, logistics, and marketing. We employ several hundred people in China, mostly in Nanjing, Beijing, and Hangzhou.

When we began building AI into our operations, the arithmetic did not favor US frontier models. Every efficiency gain would have been partly captured upstream.

Open weights solved these problems. We downloaded Qwen and Kimi, ran them on our own infrastructure, and fine-tuned them on 15 years of accumulated China e-commerce data. The models were ours to shape into customized applications, which became an AI suite called Webber, encompassing performance and analytics platforms, creative production, logistics management, and more. All of it runs on open Chinese models on our servers. The efficiency gains have been significant.

China’s framework for AI diffusion has other benefits. The government and large firms have taken a proactive stance on AI's potential social harms. Courts have found that a company cannot dismiss workers simply to replace them with AI. JD.Com has committed to retraining warehouse staff as it automates. New rules have restricted AI companion services for minors. At the World AI Conference in Shanghai, Chinese President Xi Jinping insisted that AI must serve humanity.

In line with this, earlier this year, I made a written commitment to WPIC staff that no one at the company will be made redundant because of AI.

None of this means China will avoid the disruption entirely. Some firms will make the pledge we made and quietly break it. Automation will eliminate categories of work here as everywhere else, and retraining programs are easier to announce than to deliver.

Notably, Stanford University's 2026 AI Index found that 83 percent of people in China see AI as more beneficial than harmful, compared with 39 percent of Americans. People in China seem more on board with AI, and global enterprises are increasingly leaning on China’s leading models to support their core products and business operations.

Accessible, open-weight, low-cost, focused on positive applications, and guardrails around negative externalities. China’s AI framework has significant merit.

The author of this article is Jacob Cooke, co-founder and chief executive officer of digital market services provider WPIC.

Editor: Futura Costaglione