Experts Question Kimi K3's Rapid Development Amid Allegations of Technology Theft

Washington DC, United States Jul 23, 2026 ALN: White House science advisor Michael Kratsios has raised significant concerns regarding the rapid development of Kimi K3, an open-weight large language model (LLM) developed by the Chinese company Moonshot. Kratsios alleges that the model was built by copying Anthropic’s Fable LLM while utilizing chips that are not cleared for export to China, which raises questions about the ethical and legal implications of such technology transfer.

Kratsios stated, "Large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable." His comments come amid discussions about potentially banning Chinese open-weight models due to fears of intellectual property theft, a topic that has become increasingly contentious in the AI sector. Despite the gravity of these allegations, Moonshot has not responded to inquiries regarding its training processes, and Kratsios has not provided additional details to substantiate his claims.

The concerns voiced by Kratsios align with remarks made by Treasury Secretary Scott Bessent, who indicated that there are identifiable traces of U.S. large language models within several Chinese models. Bessent’s statement underscores a growing apprehension among U.S. officials about the potential for foreign entities to exploit American innovations. However, the specific nature of these "watermarks" remains unclear, and the Treasury Department has not clarified its position on the matter.

Despite these allegations, many experts in the AI field are skeptical about the notion that the advanced capabilities exhibited by Kimi K3 can be attributed solely to distillation techniques. Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI, expressed doubts about the feasibility of distilling a model as robust as Kimi K3 so quickly after Fable's release. He noted that Fable has only been publicly available since July 1st, making it implausible to distill sufficient data, train a new model, and release it within a mere two-week timeframe.

Nathan Lambert, an AI researcher at the Allen Institute for AI, echoed Hancock's skepticism. In a recent podcast, he remarked, "I’ve been of the opinion that distillation is becoming less and less impactful over time as the Chinese models get closer to the frontier and the training regime shifts to reinforcement learning." Lambert's assertion suggests that as models evolve, the methods for replicating their capabilities also need to adapt, indicating a shift in the landscape of AI development.

Distillation, in the context of AI, refers to a process where a smaller model is trained to replicate the performance of a larger, more complex model. This process often involves querying the target model to extract information about its decision-making processes. For instance, researchers might prompt the model to explain its reasoning or use its responses to train a new model through a method known as supervised fine-tuning (SFT). However, Lambert argues that the benefits of SFT are diminishing as AI models grow increasingly sophisticated.

To effectively distill capabilities akin to those of Fable, Lambert posits that reinforcement learning techniques would likely be necessary. This approach typically involves a more complex setup, where a larger model evaluates and grades the responses of a smaller model, allowing for iterative improvements based on feedback. Such advanced techniques require substantial computational resources, often necessitating tens of millions of agents to perform adequately. Relying on a leading model’s API for such tasks could be prohibitively expensive and time-consuming, potentially negating any performance gains.

There are indications that Kimi K3 may have benefited from insights gained from earlier frontier models. Earlier this year, Anthropic publicly accused Moonshot, along with other companies like DeepSeek and MiniMax, of systematically distilling its models. Anthropic claimed to have identified millions of interactions between its models and users associated with these companies, suggesting that the queries made were not for legitimate use but rather aimed at extracting capabilities. However, Anthropic has not responded to inquiries regarding the specifics of the Fable distillation process.

While the practice of distillation has raised eyebrows, it is not exclusive to Chinese companies. Elon Musk testified earlier this year that his company, SpaceXAI, employed distillation techniques on OpenAI models to develop Grok, indicating that this practice is widespread across the industry. The distinction between distillation and the creation of synthetic datasets can often be blurred, complicating the conversation around intellectual property and ethical AI development.

Hancock further emphasized the technical capabilities of Chinese AI teams, noting that one of the founders of Moonshot was a PhD student from Carnegie Mellon University. He remarked, "These are legitimate researchers and engineers doing solid work." This perspective highlights that while there may be concerns about technology theft, it is essential to recognize the advancements made by Chinese researchers in their own right. Hancock believes that even if American models halted their progress, China would continue to advance, suggesting that the country is not merely dependent on foreign innovations.

The implications of Kratsios’ allegations extend beyond the technical aspects of AI development. They also raise significant questions about the procurement of advanced technology. Kratsios indicated that Moonshot had obtained high-performance Nvidia chips, specifically Grace Blackwell 300s, and had access to GB300-equipped servers located in Thailand. These chips are subject to export bans to China, and the existence of a black market for such technology has been noted by experts like Sam Bresnick, a research fellow at Georgetown’s Center for Security and Emerging Technology. Bresnick pointed out that in May, the founder of Supermicro, a U.S.-based server manufacturer, was indicted for smuggling advanced chips into China, highlighting the ongoing challenges in regulating the export of sensitive technologies.

Bresnick advocates for stringent "know your customer" laws for data centers globally, arguing that organizations allowing companies to conduct extensive training runs on cutting-edge hardware should implement mechanisms to track the identity and activities of those companies. This call for transparency in the AI supply chain reflects broader concerns about national security and the integrity of technological advancements.

In response to these growing concerns, President Joe Biden’s Department of Commerce proposed federal regulations in 2024 aimed at enforcing know-your-customer rules for data centers. However, progress has been slow, particularly during the Trump administration, which has hindered the development of comprehensive policies to address these issues. Exporters of advanced chips are required to ensure that their products are used for approved purposes, yet the effectiveness of these regulations in practice remains to be seen.

The ongoing debate surrounding Kimi K3 and the allegations of technology theft underscores the complexities of the global AI landscape. As competition intensifies, the intersection of innovation, ethics, and regulation will play a crucial role in shaping the future of AI development. The implications of these allegations extend beyond individual companies, as they raise fundamental questions about the protection of intellectual property, the integrity of research, and the responsibilities of nations in fostering a fair and equitable technological environment.