China’s biggest AI model is challenging American dominance

So far, the AI boom has been dominated by U.S. companies like OpenAI, Google, and Meta. In recent months, though, a new name has been popping up on benchmarking lists: Alibaba’s Qwen. Over the past few months, variants of Qwen have been topping the leaderboards of sites that measure an AI model’s performance.

“Qwen 72B is the king, and Chinese models are dominating,” Hugging Face CEO Clem Delangue wrote in June , after a Qwen-based model first rose to the top of his company’s Open LLM leaderboard .

It’s a surprising turnaround for the Chinese AI industry, which many thought was doomed by semiconductor restrictions and limitations on computing power. Qwen’s success is showing that China can compete with the world’s best AI models — raising serious questions about how long U.S. companies will continue to dominate the field. And by focusing on capabilities like language support, Qwen is breaking new ground on what an AI model can do — and who it can be built for.

Those capabilities have come as a surprise to many developers, even those working on Qwen itself. AI developer David Ng used Qwen to build the model that topped the Open LLM leaderboard. He’s built models using Meta and Google’s technology also but says Alibaba’s gave him the best results. “For some reason, it works best on the Chinese models,” he told Rest of World . “I don’t know why.”

In the short term, much of Qwen’s success comes from its unique position in the Chinese market. At launch, Alibaba claimed some 90,000 clients were using some models from Alibaba’s Tongyi Qianwen LLM series. (The name “Qwen” comes from a shortening of the term, which translates roughly to “all-encompassing knowledge.”) Most of the clients are Chinese companies that would be reluctant to form direct partnerships with U.S. companies like OpenAI or Anthropic.

Companies around the world are trying to integrate AI into their products and services, with Chinese companies being no exception. Alibaba claims that Qwen has over 2.2 million corporate users, but most of the public partnerships are still experimental. One Qwen-powered product made for Xiaomi’s mobile device division allows users to generate recipes from a photo of a dish. Qwen also powers Xiaomi’s mobile assistant, offered both on handsets and in-car systems.

Building on Qwen gives Chinese companies more stability. U.S. AI companies have cut off API access for Chinese users, while many AI sites are blocked by the central government . More importantly, though, Qwen isn’t just an available alternative: It’s proving to be just as capable as what’s available in the U.S. When it launched the most recent version of Qwen, Alibaba shared benchmark results approaching the flagship models from OpenAI and Anthropic .

Engineers can also access Alibaba’s foundational model from almost anywhere on the planet. Qwen’s fluency in major languages that lie outside most of the world’s AI training data — including low-resource languages like Burmese, Bengali, and Urdu — gives it an edge. By comparison, Meta’s open-source AI model Llama’s intended use cases cover only English.

Some in the industry see that as an extension of Alibaba’s general principle of building for the entire world, rather than prioritizing China. “I believe Qwen’s strategy exemplifies Alibaba’s early mission to ‘make it easy to do business anywhere,’” Tiezhen Wang, an engineer at Hugging Face, told Rest of World , saying the model “has the potential to revolutionize global business communication by transcending linguistic and even cultural boundaries.”

Dylan Patel, of independent research and analysis company SemiAnalysis, told Rest of World that while Qwen isn’t quite as good as GPT-4, it’s close enough to raise eyebrows. “They make claims, but, from what I’ve looked at and used, it isn’t there,” he said. But Patel says Alibaba’s model often outpaces its rivals in areas like formal mathematics and multilingual operations.

Qwen’s performance is notable given Washington’s significant trade barriers intended to slow Chinese AI development. Since 2022, the U.S. has blocked exports of Nvidia’s most advanced chips — the same chips that are powering the latest generation of AI models. Much of the equipment needed to manufacture advanced chips domestically is also blocked from export into China, so domestic tech firms like Huawei have struggled to fill the gap. As AI firms race to create ever more complex models, they need ever larger quantities of processing power — and the chip embargo means that Chinese firms run a real risk of coming up short on that processing power.

But while the chip embargo remains a factor, analysts see the recent success of Chinese AI as evidence that it hasn’t scuttled the industry entirely. “These LLM leadership boards at a minimum demonstrate how sophisticated the ecosystem in China is,” Karman Lucero, a fellow at the Paul Tsai China Center at Yale Law School, told Rest of World . “They wouldn’t be able to do this if they didn’t have a certain level of talent, if they didn’t have access to at least a certain level of technology, and if they didn’t take a certain threshold of sophistication to bear.” ▰

Sam Eifling is a writer, editor, and producer living in Brooklyn, New York.