The K3 Problem: White House Blames Distillation, but Researchers Say the Math Doesn’t Work
The White House says Moonshot copied Anthropic’s Claude in two weeks. Researchers say that’s physically impossible, and the real K3 story is far more complicated.
On July 22, 2026, Michael Kratsios, director of the White House Office of Science and Technology Policy, said his administration had information indicating Moonshot AI used Anthropic’s Fable model while developing Kimi K3. The post landed like a geopolitical flare. Within hours, Treasury Secretary Scott Bessent was on Fox Business threading in talk of sanctions, and Washington’s AI anxiety machine was running at full speed. The story was clean, satisfying, and politically useful: China stole the crown jewel, here’s the proof, consequences are coming.
There is only one problem with that story. The researchers who actually train large language models for a living took one look at the timeline and said: that’s not how any of this works.
On July 16, 2026, Moonshot AI released Kimi K3, a 2.8-trillion-parameter model that the company calls the largest open-weight AI system ever built. Kimi K3 took the number one spot in the Frontend Code Arena with 1,679 points, passing Claude Fable 5, GPT-5.6 Sol, and GLM-5.2. That jump rattled Silicon Valley. An open-weight Chinese model, available to anyone, pulling ahead of the most capable proprietary systems on a benchmark that measures real developer work. The political pressure to explain it was immense. The White House’s explanation just happened to be technically incoherent.
The Accusation, in Detail
Kratsios said Moonshot, the company behind Kimi K3, built its model by copying Anthropic’s Fable while using chips that aren’t cleared for export to China. The mechanics, as described, involved a purpose-built internal tool. “To do this they developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection,” Kratsios wrote on X. He also added a hardware dimension: Kratsios alleged that the company used GB300 servers, either newly acquired or through Thailand, to train its AI models. The GB300 servers are part of Nvidia’s Blackwell generation, which are banned from being sold to Chinese companies.
Kratsios’ post echoed comments from Treasury Secretary Scott Bessent, who said “we are finding watermarks of our U.S. large language models on many of the Chinese models, and that that’s unacceptable.” It’s not clear what those watermarks consist of, and the Treasury Department did not respond to a query. Bessent went further on sanctions: “If we see, especially that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft,” he told Fox Business.
None of this came with supporting evidence made public. Moonshot did not respond to questions about its training process, and Kratsios did not share more details about the sources of his allegations. The Chinese embassy in Washington offered a one-line rebuttal. A Chinese embassy spokesperson told Reuters the accusations were entirely unfounded.
The Fourteen-Day Problem
Here is the part that keeps AI researchers up at night, and not for the reason Kratsios intended. Some experts dispute the idea that Kimi K3 could have been developed primarily through distillation from Fable, which has only been publicly available since July 1. That leaves a window of roughly two weeks between Fable becoming publicly accessible and Kimi K3 going live. Anyone who has spent time inside a serious model training pipeline will tell you that two weeks is nothing.
“I don’t think you get a model this strong and this quickly on the heels of Fable doing strictly distillation,” Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI, told TechCrunch. “There’s just not even frankly time, right? Fable’s only been publicly available since July 1st. You can’t distill that much data, train a model, and release it in two weeks.” Hancock also pushed back against a broader narrative that Americans seem to have settled into comfortably. “In general, Americans are understating the technical expertise of these Chinese teams,” he said. “One of the founders of Moonshot was a CMU PhD student. These are legitimate researchers and engineers doing solid work.”
The technical constraints are not just about calendar time, either. The more advanced techniques also require significant infrastructure. Large reinforcement learning runs can require tens of millions of agents. Using a frontier lab’s API to do that “would be insanely expensive and potentially it would probably be a time bottleneck because these models are pretty slow and to be frank might not even give you a performance uplift.”
A Moonshot employee, Randy Xian, wrote on social media that Fable 5 became publicly accessible July 1, after a period during which U.S. export controls had suspended access, and said Kimi K3 launched July 15, calling it implausible that the company could train an entirely new frontier-scale model in that window. Nathan Lambert of the Allen Institute for AI and Elie Bakouch of Prime Intellect expressed similar skepticism about the technical timeline, and Dean Ball, head of strategic futures at OpenAI, said he did not believe distillation could account for the model’s performance.
Distillation Is Losing Its Edge — and Experts Know It
There is a second, deeper argument against the White House’s framing, and it comes from how the field has actually evolved. Nathan Lambert, an AI researcher at the Allen Institute for AI, made the case in a podcast the day after the accusations dropped.
Nathan Lambert, Allen Institute for AI:
“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. If it were the case, everyone would be easily able to catch up to a GLM or to a K3 by using its data for distillation. But we have not, or we won’t see this, from supervised fine-tuning alone.” — via TechCrunch
That is a pointed observation. If pure distillation from a stronger model were sufficient to leap to the frontier, then anyone with API access and enough money could clone the best model in the world. The reason that does not happen is that supervised fine-tuning on a stronger model’s outputs captures surface behavior, not the deep reasoning structure built through reinforcement learning. Lambert made the deeper point: if straight supervised fine-tuning on a stronger model’s outputs were the trick, everyone would be catching up to K3 the same way, and they are not. Closing the remaining gap, in his framing, takes reinforcement learning across “tens of millions of agents,” which is not something you rent through an API cheaply or quickly.
What Moonshot Actually Says It Built
While Washington was pointing fingers, Moonshot was pointing to architecture papers. Moonshot AI business head Huang Zhenxin denied the White House accusation that Kimi K3 was built by distilling Anthropic’s Fable. He credited Moon Clip, Kimi Delta Attention, and Attention Residuals for the model’s gains. These are not vague marketing claims. Kimi K3 is a sparse Mixture-of-Experts model built on two architectural updates: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), both of which change how information flows across sequence length and model depth. Kimi Delta Attention, which Moonshot introduced in the Kimi Linear paper in October 2025 and scaled up dramatically for K3, refines Gated DeltaNet by changing how the recurrent state forgets. That is a paper trail going back nine months before Fable was released.
A 2.8-trillion-parameter mixture-of-experts model is not created by collecting a few weeks of chatbot answers. Its base capabilities require enormous pretraining runs, data pipelines, routing experiments, failure recovery, evaluation, and infrastructure work. Moonshot plans to release the full weights July 27, when outside researchers can begin checking the architecture and results now available only through an API. Once those weights are public, the architecture claims can be independently verified — which is precisely the kind of scrutiny that government allegations made without supporting evidence cannot withstand.
Huang Zhenxin emphasized that “the leapfrog performance improvement of Kimi K3 relies on core, original architectural innovation at the foundational level, not the distillation or replication of any existing” model. The company claims approximately a 2.5-fold gain in scaling efficiency over Kimi K2, activating 16 of 896 experts. That kind of efficiency jump is what you get from a novel routing mechanism, not from downloading someone else’s outputs for a fortnight.
The Distillation That Actually Happened
Here is where it gets complicated, because something did happen — just not on the schedule the White House implied. Anthropic accused three Chinese AI companies of working to “illicitly extract” results from its Claude models to bolster rival products. The company said DeepSeek, MiniMax, and Moonshot violated its terms of service by generating more than 16 million exchanges in total with its Claude models using thousands of fraudulent accounts.
The February 23, 2026 disclosure was detailed. Moonshot AI ran over 3.4 million exchanges targeting agentic reasoning and tool use, coding and data analysis, computer-use agent development, and computer vision. MiniMax ran over 13 million exchanges concentrated on agentic coding and tool use, by far the largest volume of the three campaigns. Despite Anthropic’s service restrictions preventing commercial access to Claude in China, the three firms allegedly engaged commercial proxy services to sidestep Anthropic’s restrictions, enabling access to networks running tens of thousands of Claude accounts simultaneously.
None of this is about Fable, which did not launch until June 9. The February disclosure concerned earlier Claude models. It seems likely that previous frontier models might have contributed to Kimi; Anthropic publicly accused Moonshot, DeepSeek, and MiniMax of systematically distilling its models earlier this year. The honest version of the accusation, then, is that Moonshot almost certainly distilled earlier Claude models over the course of months, and that data likely contributed something to K3’s training. The White House version, pinning it all on a two-week sprint from Fable, does not hold up.
Elon Musk testified earlier this year that his company SpaceXAI distilled OpenAI models to develop Grok, and that the practice was common in the industry. The line between distillation and developing synthetic datasets can be fairly blurry. That context is conspicuously absent from official statements threatening sanctions.
Banned Chips and Southeast Asian Shell Games
The hardware side of the accusation has more structural credibility, even if the specifics remain unverified. It is hard to disentangle distillation from the second part of Kratsios’ comment: that Moonshot had obtained advanced Nvidia chips, Grace Blackwell 300s, and accessed GB300-equipped servers in Thailand. Those chips are banned from export to China, but a black market exists, according to Sam Bresnick, a research fellow at Georgetown’s Center for Security and Emerging Technology.
The existence of that black market is not speculation. Earlier in 2026, the export control enforcement apparatus caught a dramatic case. Super Micro Computer said Yih-Shyan “Wally” Liaw, a co-founder, resigned from the server maker’s board after he was indicted on allegations of smuggling equipment containing Nvidia artificial intelligence chips into China. Prosecutors alleged that the trio was involved in a scheme to send $2.5 billion worth of U.S.-made servers to China between 2024 and 2025. They used a shipping and logistics company to repackage the servers into unmarked boxes to conceal their contents before shipment to China. To deceive auditors, the men allegedly used “dummy” nonworking copies of the servers when the actual servers were on their way to China.
The Thailand routing in the Moonshot allegation follows the same playbook. Southeast Asian data centers and shell companies have become the standard bypass for chip controls, and intelligence suggesting Moonshot accessed GB300 servers there is at least operationally plausible. The problem is that Kratsios folded this into a single accusation alongside the Fable distillation claim, letting the more credible hardware allegation carry rhetorical weight for the technically dubious software one.
The Policy Gap Nobody Wants to Fill
Behind the technical dispute sits a real regulatory vacuum. The Biden administration tried to address the data-center side of the problem in 2024, proposing federal know-your-customer rules that would require data centers to report who is conducting large training runs on advanced hardware. That proposal went nowhere under the Trump administration.
Bresnick, whose research at Georgetown focuses specifically on U.S.-China technology competition, made the point directly to TechCrunch: “I am a proponent of know your customer laws for data centers across the world. If you are letting a company conduct huge training runs on your latest hardware, there needs to be a reporting mechanism for who that company is and what they’re doing.”
The reason this matters beyond a technical dispute is that the distillation narrative reshapes policy. If the story is “China copied a model in two weeks using an API,” the policy response looks like better API rate limiting and more aggressive terms-of-service enforcement. If the story is “China built genuinely novel architecture over years, supplemented by distillation of earlier models, and possibly trained on hardware smuggled through Southeast Asia,” the policy response has to be more serious, more structural, and harder to deliver in a social media post.
The White House framed the issue as a clear case of industrial espionage. Technical experts point to significant gaps in that narrative. The lack of public evidence from Kratsios, combined with the technical and timeline constraints, suggests that the story is more complex than official accusations imply.
What the Weights Will (and Won’t) Tell Us
The argument does not end until Moonshot releases the full weights on July 27, when outside researchers can begin checking the architecture and results now available only through an API. Once those weights are public, independent labs can probe for behavioral fingerprints of Anthropic’s training: unusual refusal patterns, characteristic reasoning structures, the kinds of tells that distinguish genuine architectural innovation from a model that learned from someone else’s outputs. A few prompts cannot distinguish distillation from contamination, shared evaluation data, convergent post-training, or ordinary hallucination. Strong behavioral evidence would require preregistered prompts, repeated runs, model-specific canaries, shared rare errors, and statistical controls.
The White House has access to intelligence the public does not. It is possible that Kratsios has classified evidence of Moonshot’s exact training pipeline that he cannot publish without burning sources. It is also possible that the administration saw a Chinese model beating American ones at coding, needed an explanation, and reached for the most available narrative. The researchers who have spoken on record are not saying distillation never happened. They are saying the specific claim, Fable copied in two weeks to produce K3, does not survive basic arithmetic. Those are different things, and conflating them serves neither honest policy nor genuine security.
K3 is real. It is, according to TechRepublic, an open-weight model that has reached frontier-level performance in several key benchmarks, built by a team that has spent years developing its own architecture and published the receipts. The genuinely uncomfortable truth for Washington is not that China stole the playbook. It is that China may be writing its own.
