The US State Department has ordered embassies worldwide to warn foreign governments about the risks of using AI models from DeepSeek, Moonshot AI, and MiniMax, accusing the three Chinese startups of training their systems on outputs distilled from American frontier labs. The diplomatic cable, dated Friday, April 24, instructs diplomats to raise "concerns over adversaries' extraction and distillation of U.S. A.I. models" with their counterparts, according to the document seen by Reuters.
What the cable actually says
The stated purpose is to "warn of the risks of utilizing AI models distilled from U.S. proprietary AI models, and lay the groundwork for potential follow-up and outreach by the U.S. government." A separate demarche, the document notes, has been sent to Beijing directly.
The argument leans hard on a particular framing. Distilled models "appear to perform comparably on select benchmarks at a fraction of the cost," the cable says, but "do not replicate the full performance of the original system." Worse, the campaigns "deliberately strip security protocols" and undo guardrails meant to keep models "ideologically neutral and truth-seeking." That last phrase is doing a lot of work. It is also the kind of claim that is hard to verify from the outside.
The evidence base
The cable doesn't appear out of nowhere. In February, Anthropic published an unusually detailed blog post alleging the same three labs had run roughly 24,000 fraudulent accounts and pulled more than 16 million exchanges out of Claude. Striking numbers, though they come from Anthropic's own internal forensics, not something independent researchers have audited. OpenAI sent its own memo to the House Select Committee on the Chinese Communist Party around the same time. Two days before the cable went out, OSTP Director Michael Kratsios issued a White House memo accusing China of "deliberate, industrial-scale campaigns" to distill US systems.
DeepSeek has consistently denied the framing. The company has said its V3 model trained on naturally occurring web data, not synthetic outputs from OpenAI APIs. The Chinese Embassy in Washington called the latest allegations "groundless" and "deliberate attacks on China's development and progress in the AI industry."
The timing
The cable went out the same day DeepSeek dropped a V4 preview, a 1.6 trillion parameter model optimized to run on Huawei Ascend chips rather than Nvidia hardware. Not a coincidence anyone is pretending is a coincidence. It also lands about three weeks before President Donald Trump is scheduled to meet Xi Jinping in Beijing on May 14.
What the cable actually achieves is harder to pin down. There are no new export controls or entity listings. No API restrictions either. It is a diplomatic prompt to allies and a demarche to Beijing, with a public-facing leak that lets Washington put down a marker before the summit. Several Western and Asian governments had already restricted DeepSeek for their officials over data privacy concerns, so the message isn't landing on cold ground. But DeepSeek's models remain among the most downloaded on open-source platforms, and a State Department warning is unlikely to change that part of the picture.
Trump and Xi meet in Beijing on May 14. Whether AI distillation makes the formal agenda or stays as backstage leverage will be visible from the summit communiqué.
