Washington turns model distillation into diplomacy, with DeepSeek in the crosshairs

Reuters reported late on April 26 that the U.S. State Department sent a diplomatic cable to posts around the world directing staff to raise concerns about Chinese AI firms allegedly extracting and distilling American models. That is a meaningful escalation. What had mostly looked like a fight between labs over copying, training data, and API abuse is now being turned into a formal message for foreign governments and procurement conversations.

According to Reuters, the cable was dated Friday and told diplomatic and consular staff to speak with their counterparts about the risk posed by adversaries using distilled versions of U.S. systems. It also said a separate demarche had been sent to Beijing. Distillation, in practical terms, is the practice of training a smaller or cheaper system on the outputs of a more capable one. Done lawfully, it can be a standard compression technique. Done through unauthorized access or abuse, it becomes a commercial and policy flashpoint fast.

The report matters because it connects several threads that had been moving separately. OpenAI has already warned U.S. lawmakers that DeepSeek was targeting frontier American labs to replicate capabilities, and the White House made similar accusations this week. Reuters says the State Department cable explicitly mentioned DeepSeek and also named Moonshot AI and MiniMax. That turns a private grievance into a diplomatic campaign with allies, regulators, and foreign buyers as the audience.

China rejected the accusations. Reuters quoted the Chinese Embassy in Washington calling them baseless, while DeepSeek did not immediately comment. The timing also sharpens the signal. Reuters noted that DeepSeek launched a preview of its V4 model on Friday, adapted for Huawei chip technology, a reminder that the dispute is not only about model performance but about who can build an AI stack with less dependence on U.S. hardware, software, and political approval.

The most consequential line in the report is not about one company. It is the cable’s warning that models built from unauthorized distillation can look competitive on selected benchmarks at much lower cost while shedding the original system’s safeguards. If that claim becomes the basis for procurement warnings, export controls, or partner-country guidance, then model provenance stops being a niche research argument. It becomes part of trade policy, national security screening, and the global sales pitch for AI itself.