Michael Kratsios, the White House’s top science and technology advisor, went public with a serious charge on July 22. Moonshot AI, he said, distilled Anthropic’s Fable model at massive scale to build Kimi K3. Not licensed. Not disclosed. Just taken and repackaged as their own breakthrough.
Kimi K3 hit number one on the code benchmarks a week ago. K3 is 2.8 trillion parameters, built from something that was not theirs to build from. Kratsios described it as large-scale, covert industrial distillation. Those are not policy words. Those are accusation words.
Here is what matters: we are now at the point where models are valuable enough to steal. Not the code. Not the infrastructure. The trained weights themselves. A company can take months of compute and billions in training spend from one lab and condense it into something faster, smaller, and ready to sell.
This is not theoretical IP theft. This is happening in the open, with a government official naming it directly. Moonshot has not responded with a denial. They have been quiet. That silence in the face of a White House accusation tells you something about the confidence in their own position.
The bigger picture: as models get more valuable, as the gap between frontier and efficient widens, distillation is going to become a question of industrial security. You cannot build a moat on pure capability anymore. You have to control the supply chain of trained models. That is the conversation coming.

