Conceptual diagram of AI Photo: VCG

As more US artificial intelligence (AI) companies appear to be eyeing Chinese open-weight models, Harvey, a San Francisco-based legal AI firm, has recently built its first in-house model, Harvey Tenet, on top of Moonshot AI's Kimi K3, underscoring the growing appeal of Chinese open-weight technology among US developers.

Chinese experts said that Chinese open-weight models are gaining traction not only for performance and cost reasons, but also for the flexibility of their ecosystems. As they improve in efficiency, customization and engineering capability, they are becoming harder for global developers to ignore despite mounting US political pressure. 

According to Harvey, its first post-trained open-weight model, based on Kimi K3, delivered promising initial results in both legal AI performance and cost efficiency, including what the company described as "state-of-the-art performance" on complex legal tasks.  

A string of recent cases has prompted US industry professionals to take a fresh look at Chinese open-weight models.

Harvey's pivot was "a great example of the power of open-weight models" enabling developers to post-train the model on industry-specific data or even company-specific data for more accurate and cheaper inference, AI policy researcher Simon Hedlin wrote in a post on X.

"We're still only in the very earliest stages of exploring what's possible to do with highly capable open-weight models. It's unfortunate that America is lagging behind in developing frontier open-weight models," Hedlin said in the post.

Beyond performance and price, Chinese experts said that the broader competitive edge lies in the range of models available.

One key reason US companies are likely to pay greater attention to and increasingly adopt Chinese open-weight models lies in the breadth of China's model ecosystem, Tian Feng, former dean of SenseTime's Intelligence Industry Research Institute, told the Global Times on Sunday. 

"Chinese developers have built a relatively continuous spectrum of models, ranging from smaller models such as Qwen3-4B to DeepSeek V3.2, and trillion-parameter models such as Kimi K2 and K3, while the US open-weight ecosystem offers far fewer choices outside the mid-sized segment," Tian said.

"This means companies can find models suited to different computing budgets within the same broader technology ecosystem and then deploy, fine-tune and customize them for their own needs," Tian said. 

"That is where the real advantage in being customizable comes from - not simply from one model topping a benchmark, but from the broad and continuous coverage of China's open-weight model ecosystem. In practice, professional market assessments tend to put performance, cost and business needs ahead of geopolitical rhetoric."

After Chinese AI model GLM-5.2 was used as a last resort to help Hugging Face analyze a cybersecurity incident, US industry demand for Chinese AI models has increasingly come into public view.

At the frontier scale, some US model releases above 100 billion parameters this year are built on top of Chinese AI models or leverage artifacts from Chinese labs, such as Thinking Machines' Inkling (952 billion), according to a report released by the world's largest AI open-source community Hugging Face on August 14.

The report also noted that in almost every month of 2026, the largest and most performant open model from a Chinese lab was larger than any model an American lab released. China's monthly ceiling ran between 754 billion and 2.78 trillion parameters; US models stayed under 130 billion in five of seven months, the exception being Nvidia's Nemotron 3 Ultra at 561 billion in May and June, and Inkling from Thinking Machines Lab.

Tian said that the experience of companies such as MiniMax and DeepSeek points to a common trend: if Chinese models can continue improving in performance, cost efficiency and the ability to be customized, their long-term competitiveness will depend less on any single algorithmic breakthrough than on the cumulative gains from better engineering efficiency.

Notably, this emerging trend in the US market comes as US policymakers are still casting about for ways to justify keeping Chinese open-weight models out of the US market. Chinese experts cautioned that while open adoption is already taking place, the bigger uncertainty for the industry is how difficult it will be to absorb the costs arising from political frictions.

Airbnb CEO Brian Chesky defended the company's use of Alibaba's Qwen model for its customer service chatbot in May and noted that the company has preferred the model in certain situations because it is "fast and cheap." Since rolling out its customer service agent, average resolution time dropped from nearly three hours to six seconds, the firm said last year, according to a report by Forbes.

In April, the US House committees on China and homeland security sent a letter to Airbnb, asking for clarifications on its use of Chinese AI models as part of an investigation into what they described as a Chinese campaign to "accelerate its AI capabilities by exploiting American innovation," Forbes reported.

"The tension between US politicians' so-called national security restrictions and companies' practical needs is unlikely to disappear, and US firms are already bearing the cost," Tian said. "But the adoption of technology will not simply wait for geopolitical concerns to fade."