China’s commerce ministry fired back Monday at fresh American charges that its AI developers had stolen advanced model capabilities through a process known as distillation. The sharp rebuke accused Washington of “AI hegemonism” and vowed to take all necessary measures if Chinese interests suffer harm. The exchange marks another flashpoint in the intensifying contest for dominance in artificial intelligence.
Distillation allows a smaller model to learn from the outputs of a larger, more powerful one. It compresses performance into efficient systems that run on less hardware. American officials now frame large-scale use of the technique by Chinese firms as outright theft. They point to cases where companies allegedly created thousands of fake accounts to query US models like Anthropic’s Claude or its newer Fable system at industrial volumes.
Anthropic laid out the pattern in a February statement. It identified three Chinese labs — DeepSeek, Moonshot and MiniMax — running coordinated campaigns. These efforts generated more than 16 million exchanges through roughly 24,000 fraudulent accounts. The goal: siphon reasoning, coding and tool-use skills that cost hundreds of millions to develop. Anthropic called the activity systematic, not casual misuse. (Anthropic)
The White House amplified those concerns last week. Michael Kratsios, director of the Office of Science and Technology Policy, singled out Moonshot AI. He claimed the Beijing-based creator of the Kimi chatbot had distilled Anthropic’s Fable model to build its K3 release. Kratsios described an internal platform designed for large-scale extraction. It rotated access methods to dodge detection. Some operations reportedly routed through Thailand to mask origins. (CNBC)
But Beijing rejects the narrative. Its commerce ministry called the US allegations groundless. They lack facts or legal foundation, the statement said. And they apply double standards. The ministry turned the tables. Many American AI enterprises have distilled Chinese models during their own research and training, it asserted. The claim lands as a direct mirror to Washington’s complaints.
“For any action that causes substantive harm to Chinese interests, China will take all necessary measures to firmly safeguard its legitimate rights and interests,” the spokesperson declared. The language carries weight. It signals readiness for retaliation. Sanctions, export curbs or other steps could follow. (Reuters via Yahoo News)
The Technical and Commercial Stakes
Distillation isn’t new. Researchers have used it for years to shrink models and cut costs. Yet scale changes everything. When applied across millions of queries against frontier systems, it transfers sophisticated behaviors without the original training data or billion-dollar compute clusters. Chinese developers gain speed. They release competitive offerings at fractions of the price.
Moonshot’s Kimi K3 drew immediate attention upon release. Users and evaluators found it matched or approached the performance of top US models from Anthropic and OpenAI in several tasks. That rapid parity fueled suspicions. Independent tests on platforms like Fireworks AI highlighted strong results in reasoning and coding. The episode crystallized fears that distillation erodes the US lead built on massive investments and chip restrictions.
Those export controls remain a pillar of American strategy. Limits on advanced Nvidia GPUs aim to starve Chinese labs of the hardware needed for frontier training. Distillation offers a workaround. It lets developers bootstrap smaller models using API access or other means. Enforcement grows tricky. Queries can hide behind proxies, fake credentials and distributed accounts. Anthropic and others have improved detection. Still, determined actors adapt.
The Foreign Affairs essay from late May captured the tension. Chinese firms extract capabilities from US systems while American companies face contractual bans on similar practices. Terms of service prohibit using outputs to train rivals. Chinese entities operate under fewer constraints, at least in practice. The result distorts markets. Open-weight models distilled from US frontiers flood global use. They run cheaply on modest hardware. Many flow back into the United States. (Foreign Affairs)
US industry voices split. Some executives worry new rules could backfire. Over twenty companies, including Nvidia, Microsoft, Meta and Palantir, signed a letter this month. It urged policymakers against premature curbs on open-weight models. Such moves might stifle competition and push innovation abroad, they argued. The letter highlights a core friction. Distillation sits at the intersection of legitimate efficiency gains and alleged intellectual property loss.
Legislation has advanced in Congress. The Deterring American AI Model Theft Act of 2026 passed the House Foreign Affairs Committee unanimously in April. It defines unauthorized extraction as a national security threat. The bill directs identification, punishment and deterrence efforts. The administration’s National Security and Technology Memorandum 4, issued around the same time, reached similar conclusions. It spotlighted deliberate campaigns by Chinese entities. (CNAS)
Yet China shows no signs of retreat. Its statement Monday flipped the script. By accusing US firms of distilling Chinese models, Beijing paints a picture of mutual behavior. The tactic seeks to undermine American moral authority on the issue. It also courts global audiences wary of US dominance in technology standards. Recent X discussions reflect the divide. Posts debate whether enforcement will slow China or simply drive more sophisticated evasion. One analyst noted that controls have kept China’s share of leading AI compute to just 1 to 4 percent. Distillation narrows that gap without new chips.
The episode builds on months of escalation. Anthropic’s initial disclosures in February triggered wider scrutiny. Google has observed similar patterns without naming specific actors. OpenAI and others report parallel incidents. The technique’s effectiveness became clear with models like DeepSeek’s offerings, which rival Western performance in benchmarks while claiming lower resource use.
So the stakes extend beyond any single lab. Control over model weights, training methods and distribution channels will shape AI adoption worldwide. Cheaper, distilled systems from China could become defaults on billions of devices, especially in price-sensitive markets. American frontier labs risk subsidizing competitors. They bear the expense of breakthroughs that get repackaged quickly overseas.
Officials in Washington draw distinctions. Kratsios stated that legitimate distillation for smaller, efficient models plays a vital role in open innovation. But covert, industrial-scale efforts aimed at stealing proprietary technology cross the line. Sanctions and entity list additions sit on the table, Treasury Secretary Scott Bessent signaled on X. The administration blends support for open models with targeted enforcement against bad actors.
Beijing’s response carries familiar rhetoric. It rejects the theft label. It portrays US actions as hegemonic overreach designed to maintain technological supremacy. The ministry’s reference to American distillation of Chinese models lacks specific examples or evidence in public statements. Still, the symmetry aims to neutralize criticism.
Industry watchers expect further measures. Enhanced monitoring of API traffic. Stricter identity verification. Potential limits on query volumes for certain regions. Yet technical fixes alone may not suffice. The cat-and-mouse dynamic favors determined state-backed players with resources to iterate.
And the pace accelerates. MiniMax, one of the named labs, claims its latest model assisted in its own training. That self-improvement loop signals where the field heads. Each distilled generation locks in gains. It feeds into domestic ecosystems that operate with state support and alternative hardware paths.
The Information first reported elements of Beijing’s briefing Monday, drawing from commerce ministry statements. Subsequent coverage by Reuters detailed the “all necessary measures” pledge. Eagle Intelligence Reports noted the deliberate timing as a counter to US pressure. These accounts align on the core facts: accusations met with accusations, threats met with threats. (The Information) (Eagle Intelligence Reports)
Washington has drawn a clearer line. Distillation moves from technical footnote to policy flashpoint. How both sides enforce their positions will influence who sets the terms for AI’s next phase. Short-term friction seems certain. Longer-term outcomes depend on execution, adaptation and the relentless march of model capabilities.
Chinese developers continue releasing strong open models. US labs push frontier boundaries behind closed doors. The public battle over methods like distillation reveals deeper contests over data, compute, talent and rules. Neither side appears ready to yield ground.
WebProNews is an iEntry Publication