What Chinese AI model Kimi’s success says about the next phase of US-China AI race

A similar sequence of events seems to be playing out now. Chinese company Moonshot AI released its Kimi K3 model earlier this month, said to rival models from US-based companies OpenAI and Anthropic. It boasts features including coding, financial consulting, deep research, and video editing, and as demand surged, Moonshot AI paused sign-ups.

Alongside its rise, Anthropic has accused Moonshot AI of illicitly extracting the capabilities of its Claude model, with one US government official describing it as “an assault on every economy that prizes the entrepreneur, rewards private capital, and relies on fair and honest competition.” On Monday (July 27), China’s Commerce Ministry accused the US of “AI hegemonism” in response.

“Markets react strongly when a Chinese release challenges assumptions about permanent US technological leadership. Another sell-off is certainly possible,” said Lizzi C Lee, a Fellow on Chinese Economy at the US-based Asia Society Policy Institute’s Center for China Analysis. But more than just being good, the Kimi story shows that Chinese models are now becoming more ubiquitous, she said.

What Kimi K3 represents

Moonshot AI was founded in 2023 by Yang Zhilin, a 34-year-old entrepreneur who graduated from Beijing’s Tsinghua University and earned his PhD at Carnegie Mellon University in the US. He worked with American tech companies before returning to China.

Kimi K3 has been billed as the world’s largest “open-weight” AI system, with 2.8 trillion “parameters”. In the AI context, parameters are the settings that determine how systems respond to user prompts, based on the large amounts of data they have been trained on. Parameters are assigned “weights” or numerical values that indicate their relative significance, which guide the model’s response.

Lee said that a massive parameter count indicates large investment in compute (software and hardware used for training and running AI systems), data and engineering. But the model’s capability also depends on factors like data quality.

Poe Zhao, a China Tech Analyst and Founder of the online publication Hello China Tech, explained that “An open-weight model allows developers to download its parameters, run it locally and customise it. Open source usually implies broader disclosure, including enough code and training information to understand and modify how the model was produced.”

A worker stands at the booth for Chinese AI Startup Moonshot’s Kimi K3 during the World AI Conference in Shanghai, China on July 17, 2026. (AP Photo/Ng Han Guan)

Conceptually, “open-source” is not limited to AI. It simply means that the underlying code behind software and programs is publicly shared so anyone can use and modify it to create their own programs. Openness is often seen as a way to democratise technology, but has also drawn counterarguments, especially on malicious actors manipulating code and posing security risks.

The open-close debate

US companies have generally not followed the open model path, while many Chinese AI labs have made their models open-weight. That decision is not just a technical one.

Zhao said that for Chinese AI labs, open weights are mainly a “distribution strategy”. “Leading US companies already have global cloud platforms and enterprise customers. Chinese labs can use open weights to reach developers more quickly, build an ecosystem” and generate demand for their products, he said.

Chim Lee, a Senior Analyst focusing on China at the Economist Intelligence Unit, told The Indian Express that Chinese developers often face “chip supply constraints, given Western export restrictions and domestic production bottlenecks.” Semiconductors or specialised chips are crucial hardware required to run AI models, and the US government has sought to restrict advanced variants made by companies like Nvidia from reaching China.

Cross-border restrictions as part of the larger geopolitical rivalry mean that Chinese labs “may not have the capacity to provide commercial access to a closed model,” Chim Lee said. “On the other hand, low-cost open-weight models are attractive to developers, the community of which often innovate and build on top of them.”

What makes the Kimi case unusual, he said, “is that it is one of the most popular models trying a custom licence approach — open-weight for most, but requiring large companies to strike a commercial agreement with Moonshot. This is an attempt at monetisation.”

Lizzi Lee said that the approach also fits China’s broader industrial strategy: “Beijing wants AI to spread quickly across manufacturing, cloud services, local governments and smaller companies. Open weights help accelerate that diffusion.” “There is a diplomatic angle, too. China increasingly presents open models as part of international technological cooperation,” she said. China has projected its models’ choice as part of a more equitable approach.

Just this month, the Chinese government spearheaded the creation of a new AI governance body called the World AI Cooperation Organisation, with President Xi Jinping stating that “AI development should not be a solo performance by a single country, but a symphony of international cooperation”.

Future of the AI race

The Kimi K3 comes at a time of several overlapping, AI-related debates in the US, touching upon tech access, national security, and the dominance of a few players.

In a recent open letter, supported by the likes of Microsoft’s Satya Nadella, Nvidia’s Jensen Huang and OpenAI, industry giants called for shifting towards open-weight models. “Software developed by the open-source community now supports most of the internet and underlies systems used by the world’s largest technology companies, as well as the U.S. military and federal agencies… The United States now faces a similar choice with artificial intelligence,” it said.

Zhao said that the letter is an “important industry signal, although it is too early to call it a major shift… A real shift would require US companies to release competitive open-weight models consistently.” While companies behind AI infrastructure (like Nvidia) have until now advocated for open-weight models, presumably to spread adoption and the demand for their products, companies that created proprietary models (like Anthropic) have had reservations.

Anthropic CEO Dario Amodei clarified that they did not sign the letter due to safety concerns, even naming China and the Chinese Communist Party (CCP) several times in a post. Within China, the government is depicting the debate as a sign of its AI ascendancy. An article in the CCP mouthpiece Global Times recently said, “The fact that Silicon Valley is now forced to argue over open versus closed publicly is itself proof that China has moved to the center of the game.”

Another debate concerns Anthropic’s allegations against Moonshot AI and DeepSeek. In February, it accused them of “distillation” — where AI models are made to interact with stronger, existing models to mimic their responses and learn from them.

While distillation is not illegal, Anthropic flagged large-scale distillation that creates new models potentially lacking in “necessary safeguards” — where AI might be used to build bioweapons or carry out malicious cyber activities. Experts told The Indian Express that these charges are serious, but remain difficult to verify in the absence of clear legal definitions so far.

Lizzi Lee said that the story has “clear parallels with earlier disputes involving semiconductors, electric vehicles and other industries. Rapid Chinese technological catch-up has often been followed by arguments over technology transfer and intellectual property.”

Continued arguments have not stopped the US from imposing export controls and other restrictions on Chinese AI. The government is also reportedly investigating whether Moonshot AI illegally accessed advanced chips to train its models. Axios reported that there may even be plans to ban Chinese open-weight AI models altogether.

One reason for this heightened activity may be increasing adoption of Chinese AI within the US. For price-conscious users, like smaller firms, using cheaper models is efficient and open weights allow for need-based customisation. “Large companies still face concerns about security, data governance and regulatory uncertainty,” Zhao said.

With every new advanced Chinese model, the gap between the countries is narrowing, but with caveats. Zhao argued that export controls on chips, for instance, have also encouraged China’s investment in efficiency and domestic alternatives. However, “K3 does not prove that export controls have failed. It shows that progress in AI models is influenced by more than access to the most advanced chips,” he said.

Parity with the US may still be far off, he said, noting the major advantages that US companies retain in compute, capital, global distribution and advanced chips. “Kimi K3’s wider significance is that a model can remain behind the absolute frontier and still create global competitive pressure if it is capable, affordable and openly available,” he said.

The rise of Chinese AI companies could also offer other countries “more choice, lower costs and greater ability to deploy models locally.” This, in turn, could give Chinese companies more influence over technical standards, further intensifying what is becoming a two-player race for global AI dominance.