Concerns Over China’s Moonshot AI and Latest K3 Model

China’s Moonshot AI has reportedly contributed to Anthropic’s latest K3 release, built on the advanced Fable large-scale language model. This was revealed by Michael Kratsios, a senior technology official in the White House.

In a post on X, he indicated that Moonshot AI has essentially distilled Anthropic’s Fable to create the K3 model. He emphasized that the large-scale, covert efforts to distill American technology and research is entirely unacceptable.

Furthermore, Kratsios mentioned that Moonshot AI has procured a potent Nvidia AI chip, known as the GB300. This chip is being utilized in Thailand, possibly for training their own models. A senior figure from the Trump administration had previously flagged concerns about another Chinese AI startup, DeepSeek, which had trained its models using Nvidia’s advanced AI chips, raising potential breaches of U.S. export regulations.

As of now, Moonshot AI, along with the Chinese Embassy in Washington, has not addressed these allegations.

Earlier this year, the U.S. State Department initiated a global campaign to highlight the ongoing attempts by Chinese companies, including Moonshot AI, to appropriate intellectual property from American AI laboratories, based on diplomatic documents reviewed by Reuters.

The term “distillation” refers to the practice of using outputs from larger, more sophisticated AI models to train smaller ones in a bid to lower costs associated with developing powerful new AI tools.

In February, reports indicated that OpenAI alerted lawmakers regarding DeepSeek’s intentions to replicate its models, posing a threat to leading AI companies in the U.S.

An April cable from the State Department articulated that models arising from unauthorized distillation processes could allow foreign competitors to release products that, while appearing effective in certain benchmarks, would not match the full capabilities of the original systems.

Recently, Moonshot AI unveiled the KimiK3 model, boasting an impressive 2.8 trillion parameters. It claims to be the largest open-weight AI system globally, performing at levels comparable to Anthropic’s Frontier Fable model.

This announcement follows the abrupt withdrawal of Anthropic’s Fables and Myths model by the U.S. government due to security issues, further underscoring the rapid development of China’s AI landscape in relation to American innovations.