Moonshot AI Accused of Covertly Distilling Anthropic Models for K3

Recently, US officials accused Moonshot AI of conducting large-scale, covert distillation from US frontier models (such as Anthropic) during the development of the K3 model. According to the allegations, Moonshot AI built a complex internal platform specifically for this purpose and evaded detection by switching between multiple access methods. Furthermore, the report noted that Moonshot AI purchased GB300 servers and deployed them in Thailand. This incident has quickly sparked widespread attention and discussion within the AI community.

Reactions and Controversy

The accusations have drawn sharply divided reactions from the community. On one hand, some practitioners like @daniel_mac8 and @Teknium focused on the allegations themselves. @daniel_mac8 pointed out that the most shocking aspect is the extreme complexity and sophistication of Moonshot AI's detection-evasion tactics; @Teknium, while relaying the accusations, brought up another controversy, alleging that Anthropic also engaged in scraping content from multiple platforms during the development of Fable. On the other hand, @MaziyarPanahi strongly refuted the claims of "large-scale covert industrial distillation." He argued that such distillation simply did not occur, emphasizing that even if provided with the logits of top-tier models directly, most European AI companies would struggle to build an excellent 3B model. This was intended to defend the genuine R&D capabilities of Moonshot AI (and models like Qwen2.5).

2026-07-22 ~ 2026-07-23 · 9 related posts