Moonshot Denies Distilling Fable and Credits K3 Gains to Its Own Architecture

Moonshot AI business head Huang Zhenxin denied on Tuesday, July 21, the White House’s accusation that Moonshot distilled Anthropic’s Fable to build Kimi K3, Asia Times reported. He credited Moon Clip, Kimi Delta Attention and Attention Residuals for the model’s gains. Moonshot says it plans to release the full weights July 27, when outside researchers can begin checking the architecture and results now available only through an API.

Moonshot’s K3 release materials

Asia Times reported that Huang said Moon Clip doubles training efficiency while cutting computing costs in half. He attributed a tenfold context-window expansion to Kimi Delta Attention and a 25 percent increase in reasoning speed to Attention Residuals. Those performance figures are Moonshot’s own claims.

The company’s Hugging Face release page describes K3 as the world’s first open 3T-class model and identifies Kimi Delta Attention and Attention Residuals as architectural changes. It lists July 27 as the expected date for the open weights.

Independent AI researcher Nathan Lambert’s July 21 analysis quoted Moonshot’s launch blog, which says K3 is “effectively activating 16 out of 896 experts when paired with a Stable LatentMoE framework.” The post says Kimi Delta Attention and Attention Residuals are designed to change how information moves across sequence length and model depth. It claims the structural changes, together with refined training and data recipes, yield about a 2.5-fold gain in scaling efficiency compared with Kimi K2. The researcher traced Kimi Delta Attention to the Kimi Linear paper, compared it with Gated DeltaNet as used for the Olmo Hybrid model and noted a related architecture in Alibaba’s latest Qwen models. He said Gated Delta Networks emerged in late 2024 from ideas in Mamba and had reached frontier models by mid-2026.

K3’s July leaderboard placements

Lambert dated Moonshot’s K3 preview to July 16 and recorded the model in second place on the Vals AI index, third on Artificial Analysis’s Intelligence Index and first in Frontend Code Arena. Claude Fable and GPT-5.6 Sol Max ranked above K3 in the Artificial Analysis snapshot, while K3 was cheaper. His assessment was that any contribution from adversarial distillation was “at most to a relatively small degree.”

Arena placed K3 first in its Frontend Code Arena with 1,679 points, Tom’s Hardware reported, ahead of Claude Fable 5. The July 16 result represented a 17-place jump from Kimi K2.6 and put K3 first in six of the arena’s seven frontend domains. The outlet cautioned that every benchmarked result came from API access and could not be verified before the weights became available.

Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI, challenged the White House’s timeline in comments to TechCrunch. “Fable’s only been publicly available since July 1st,” he said. “You can’t distill that much data, train a model, and release it in two weeks.”

The White House’s July 22 post

CyberScoop reported that White House science adviser Michael Kratsios said the administration had information that Moonshot distilled Anthropic’s Fable and built a platform to conduct large-scale distillation against U.S. models. The report noted that he provided no details about how the government obtained that information.

In a reply on X, Anthropic public-policy head Sarah Heck called illicit adversarial distillation “IP theft and industrial espionage that supports adversary military and intelligence capabilities.”

A Chinese embassy spokesperson in Washington told Reuters the accusations were “entirely unfounded.” The spokesperson urged U.S. officials to “respect the facts, discard prejudice and stop smearing and discrediting China’s achievements” in artificial intelligence.

Moonshot’s expected July 27 release

Publishing the weights will permit independent inspection of K3’s architecture and attempts to reproduce the scores. Moonshot’s release page does not mention a training-data disclosure.

White House Accuses Moonshot of Distilling Fable and Using Banned Nvidia Chips

Michael Kratsios, director of the White House Office of Science and Technology Policy, accused Moonshot AI on Wednesday of distilling Anthropic’s Fable to develop Kimi K3, the 2.8-trillion-parameter m

The Implicator

Jensen Huang Defends Chinese AI Models Hours After Bessent Sanctions Threat

Nvidia CEO Jensen Huang told Axios on Tuesday that American companies should "absolutely" be allowed to use Chinese AI models. The remarks came hours after Treasury Secretary Scott Bessent threatened

The Implicator

China Considers Adding AI Model Weights and Chip Designs to Export List

China's Ministry of Commerce is consulting Alibaba, ByteDance, Zhipu and other domestic AI and chip companies on possible controls affecting model weights, key data used for training and chip designs,

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai