What is China's Kimi K3 and why is the US so rattled by it?

Hong Kong -- Remember DeepSeek? The Chinese artificial intelligence startup that rattled the US tech industry with its model that nearly matched leading rivals, trained at a fraction of the cost? China may have done it again.

The latest Chinese AI model making waves was developed by a lesser-known startup called Moonshot AI. In just two days after launch, Moonshot suspended new subscriptions after demand overwhelmed its computing capacity.

The unexpected enthusiasm for Moonshot's Kimi K3 model stems from its ability to match -- and in some cases outperform -- its most advanced US competitors, such as the latest iterations of OpenAI's GPT and Anthropic's Claude, for free.

Now its rise is intensifying an already contentious race between the US and China to dominate cutting-edge AI. Washington has ramped up export restrictions against China in recent years to slow the country's development of AI and chips that could power its military.

After Kimi K3 was released, US officials and Anthropic have again accused Moonshot of distilling American models, a technique in which one AI model is trained using outputs from another. Moonshot has denied the allegations.

Still, the fears spreading among US companies and government officials underscore the stakes of the AI rivalry, and its far-reaching implications for economy, defense and national security.

More broadly, Kimi K3's capabilities are renewing questions about the effectiveness of Washington's tech export curbs. The new model also challenges an industry belief that leading-edge AI requires heavy spending on data centers and the advanced chips that the US has restricted access to.

Unlike most American proprietary models, Kimi K3 was released as an open-weight model. That allows developers to download it, modify it, and run it on their own servers for free, positioning it for easier adoption and a wider market globally.

Constrained by limited access to AI hardware and investment compared with American firms, China has embraced so-called open models, in contrast to US rivals that mostly sit behind paid subscriptions.

Many organisations value the ability to self-host open models, allowing them to retain the weight and control over sensitive data instead of relying solely on closed-model providers, said Chelsey Tam, senior equity analyst at financial services firm Morningstar.

But analysts say Moonshot's achievement should be viewed through a different lens. DeepSeek challenged the conventional wisdom that developing an advanced AI model requires massive financial investment and vast computing power, and offered competitive pricing. But K3's discount to frontier models is much smaller than R1's was at launch, suggesting its cost advantage may be less significant than DeepSeek's, according to Ivan Su, director at Morningstar.

What K3 demonstrates is that Chinese companies are continuing to build on the success of DeepSeek, and close the gap between Chinese and US AI models, said Poe Zhao, analyst and founder of the newsletter Hello China Tech.

"Its broader significance is that DeepSeek increasingly looks like the beginning of a pattern," he said. "DeepSeek was a surprise. K3 is evidence that China's progress is becoming more repeatable."

Since DeepSeek, Chinese AI firms like Moonshot, Alibaba and Z.ai are getting more attention for each new model, with the pace of the rollouts speeding up.

With K3, Wei Sun, principal AI analyst at market analysis firm Counterpoint Research, said the general gap between Chinese and American models has narrowed to three to six months, though it varies significantly by task.

Still, other analysts have expressed reservations toward K3's true capabilities. "While K3 constitutes progress, we'd hesitate to ascribe it near-parity with American frontier models, such as Fable 5, in real-world tasks," said Malik Ahmed Khan, senior equity analyst at Morningstar.

For American tech companies, free and comparable Chinese models like K3 threaten to undermine their own products. Leading US AI firms Anthropic and OpenAI have long accused Chinese labs, including Moonshot, DeepSeek and Alibaba, of extracting capabilities from their models. Those claims have intensified with the arrival of K3.

Huang Zhenxin, Moonshot AI's head of enterprise business, denied the accusations of distillation.

But that has not stopped new allegations from Washington. "We have information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model," said Michael Kratsios, director of the White House Office of Science and Technology Policy. "Large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable."

Sarah Heck, Anthropic's policy head, echoed Kratsios in an X post that accused Moonshot of engaging in "IP theft and industrial espionage," calling it a "national challenge that creates serious national security risks."

Treasury Secretary Scott Bessent also reiterated Wednesday that the US could sanction Chinese firms over such IP theft. "When PRC firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table," he said on X.

Most analysts, however, found the distillation accusation against Moonshot unlikely.

Sun of Counterpoint said Moonshot would not have had enough time to train its AI model with its rivals' latest models, which had limited availability before K3's launch. And distillation alone would not account for Kimi K3's outperformance, she added.

"The timeline is unconvincing," she said.