Moonshot AI, the Beijing lab founded in 2023, released a 2.7-trillion-parameter open-weight model called Kimi K3 on July 16, staged for maximum effect alongside President Xi Jinping’s speech that same day at the World Artificial Intelligence Conference in Shanghai. By Friday the Nasdaq was off roughly one percent, chip stocks led by Nvidia had been sold, and a policy debate the Trump administration preferred to defer was back on the table.

Moonshot billed K3 as its most powerful open-source coding system to date, conceding it trails Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol on overall capability while claiming it “consistently” outperformed remaining tested systems, including Claude Opus 4.8 and GPT 5.5. Independent evaluations from Arena.ai and Vals AI, cited by TechCrunch, put K3 in the same conversation as flagship frontier systems. The price is where the release does its real work: $15 per million output tokens against $50 for Anthropic’s Fable, with DeepSeek V4 still undercutting everyone at $0.87.

The commercial pull is already visible on the U.S. side. Cursor built its Composer 2 coding agent on earlier Kimi models, and DoorDash CTO Andy Fang wrote in early July that his team delegates “lower-level work to Kimi K2.6.” Moonshot’s business is scaling in step: a $2 billion round in May at a valuation above $20 billion, with its financial advisor telling investors annual recurring revenue had passed $200 million.

Which is why the second story is a distillation story. Screenshots circulated over the weekend of K3 identifying itself as “Claude, an AI assistant made by Anthropic.” Hamish Low, a research associate at the Institute for AI Policy and Strategy, called it “pretty overwhelming likely they are distilling from leading U.S. models,” pointing to K3’s strength in coding but not cybersecurity as the giveaway. OpenAI’s Dean Ball pushed back, writing that he didn’t believe K3’s performance “can be explained away by distillation or anything like that.” Anthropic went public with the same complaint against Moonshot, DeepSeek, MiniMax, z.ai and Alibaba roughly six months ago.

The administration’s response arrived through Treasury. Asked about the model by Maria Bartiromo on Fox Business, Secretary Scott Bessent said Washington “supports open-source models, but what we do not support is IP theft,” and warned that Washington retained “the ability to sanction” foreign laboratories found to have trained on U.S. systems.

That framing echoes the export-control logic that followed Huawei’s 2019 designation: a bright-line IP grievance used to justify a much broader restrictionist turn. Bessent’s sentence is the tell. Open-source, yes; open-source from Beijing at a third of the price, apparently a sanctions question.

Sources