China now leads the U.S. in this key part of the AI race
Free artificial intelligence technology released by Chinese tech companies appears to be more powerful and popular than that developed by American rivals.
Oct. 13, 2025
by
Kevin Schaul
The artificial intelligence boom started in the United States, but companies from China are quietly outcompeting their U.S. rivals when it comes to AI technology that anyone can freely use and build upon, according to a Washington Post analysis of publicly available data.
Last year, the best freely available or “open” AI models were largely made in the United States. Now, they are all made in China.
American companies are widely seen as offering the most powerful proprietary AI tools, such as OpenAI’s ChatGPT and Google’s Gemini chatbots. But by openly sharing AI software, Chinese firms could have a major influence over the trajectory of technology.
Entrepreneurs and researchers often use open software as a cheap and adaptable way to develop and launch new ideas. Trump administration officials have claimed the
U.S. must lead in open AI technology to maintain its edge in AI.
Source: LMArena
Open models from Chinese firms such as e-commerce giant Alibaba are rated higher than those from American companies such as OpenAI and Meta on LMArena, a site that uses blind tests to discover which AI outputs users prefer.
Chinese companies such as Alibaba, the maker of popular Qwen open models, are more prolific than other AI developers, said Irene Solaiman, chief policy officer at Hugging Face, a popular site used by AI developers and researchers to share models and datasets. They are “shipping frequently and shipping well, which is also how you build your user base,” she said.
In previous tech shifts such as the rise of the internet and smartphones, free and
open source technology that others can modify has played a crucial role. Much of it originated with U.S. firms, helping establish Silicon Valley’s global dominance.
For some time, U.S.-based Meta offered the world’s best open AI model. Meta founder Mark Zuckerberg argued in an
essay last year that the world would be better off if AI companies shared their best models and research for others to use freely. He pledged his company would do so with its
Llama series of releases.
But Chinese start-up DeepSeek shocked the tech world
in January with its strong open-weight model — the industry term for AI models that can be downloaded and run on any computer. On LMArena, it quickly displaced Meta’s Llama models.
One way to track the influence of a piece of open software is by its popularity with hackers, researchers and entrepreneurs working on new ideas. Hugging Face provides the leading platform for sharing open-weight AI models, where developers click “like” on projects they support.
By this measure, China’s DeepSeek is on top, with twice as many likes (12.8k) as Llama’s top model (6.3k). OpenAI ranks fifth (4.0k).
No metric is perfect, and others look rosier for the U.S.
Artificial Analysis, a company that tests AI models on such tasks as trivia, math and coding, placed an open model released by ChatGPT maker OpenAI in August a tick ahead of DeepSeek.
OpenAI’s release transformed the company into an American champion for open source alongside Meta, said Solaiman of Hugging Face. But in addition to being more prolific, Chinese competitors are also highly competitive in AI for other use cases, she said, releasing state-of-the-art open software for generating images and videos.
Meta declined to comment. OpenAI did not respond to a request for comment. The Washington Post has a content partnership with OpenAI.
Giving away software might sound counterintuitive for profit-seeking companies, but they can benefit when projects are successful enough to create an ecosystem or become an industry standard. Most smartphones around the world run Google’s Android operating system, which helps drive more users to the company’s search, email and other services.
Releasing AI models free can be more complicated than for other types of technology. Traditional software mostly behaves predictably. AI products can do complex things with
serious consequences, many of which are not well understood.
That can be an argument for keeping the technology closed or opening AI models up, depending on a person’s perspective.
Leading American AI companies such as OpenAI and Google have generally chosen to keep their best technology closed and make any open models they release less powerful. Those decisions could shape whose technology the world builds on, and the values built into future AI products and services.
AI makers choose which data guides the technology, which behaviors to reinforce and which outputs are unsafe. If AI becomes foundational to many other technologies and products, the values of the winning models could flow outward.
The popularity of China’s open-weight models concerns some in the U.S., who fear they will spread the values of the country’s government. The Trump administration’s
AI strategy urges development of open AI technology “founded on American values” that become global standards with “geostrategic value.”
Other AI experts suggest that focusing on U.S.-China competition can be misleading. The nonprofit AI Now Institute
said in a July report that the concept of an “AI arms race” with national security at stake has allowed tech companies to win tax exemptions and avoid scrutiny.
The eagerness of Chinese companies to share their best AI models — and the hesitance of U.S. firms to do the same — raises the question: Will the best open models always be made in China?
Some American executives, investors and academics think they have an answer. The
ATOM Project, short for American Truly Open Models, aims to create a U.S.-based AI lab to develop and release open-source AI models that compete with the world’s best.
Without initiatives like that, U.S. firms may be destined to remain on their current course. One year after his original essay arguing for building open models, Zuckerberg shared a new
essay. The company would need to be “careful about what we choose to open source,” he wrote. Meta might keep its next model for itself.
