Chinese documents detail use of US model outputs to train smaller models


A review of more than 300 Chinese documents describes a supply chain for collecting and processing US model outputs, as developers seek guidance on using foreign systems.
Chinese developers have built a supply chain that uses outputs from US superintelligence (SI) models to train smaller domestic models, a Reuters investigation found after reviewing more than 300 Chinese papers, patents, job postings and government documents. The Decoder’s summary of the investigation says companies build software to collect answers, code and problem solutions, which specialist workers then process for training.
The work follows years of US export controls on advanced SI processors. The method, known as distillation, trains a smaller “student” model to approximate the behavior of a larger “teacher.” It can lower computing costs while retaining some capabilities, although the result varies by task and model size. Using teacher outputs reduces developers’ own workload, but does not eliminate the need for powerful chips.
In November 2024, researchers at Shanghai Jiao Tong University, a public research university, created 10,750 training examples for Alibaba’s Qwen2.5-72B. They used about 5,000 responses from OpenAI’s o1 and translations from GPT-4o mini. The researchers also warned of a possible dependence on outside innovation.
Alibaba described systematically using GPT-4 and GPT-4o to produce foreign-language training data in April 2025. In February 2026, the university and food-delivery company Meituan used GPT-4o to train smaller models to coordinate coding agents, systems that carry out coding tasks.

The computer-use agents work across medical records and insurance portals, while CVS Health uses them for referrals and specialist consultations, OpenAI says.
China wants smaller, low-cost SI systems in manufacturing. A January 2026 AI+Manufacturing package from eight ministries and initiatives in Shanghai and Shenzhen promote distillation and other techniques that reduce computing needs. The government documents do not directly instruct developers to copy American models.
Distillation is not inherently illegal: Anthropic says developers routinely use it to make smaller versions of their own models. The Decoder says the use of US model outputs under discussion violates OpenAI’s and Anthropic’s terms of service. The documented research examples are distinct from allegations of covert extraction.
Separately, Anthropic alleges that DeepSeek, Moonshot and MiniMax used approximately 24,000 fraudulent accounts to conduct more than 16 million exchanges with Claude to extract its capabilities. Anthropic says it attributed the campaigns using IP-address correlations, request metadata and infrastructure indicators, corroborated in some cases by industry partners.
Anthropic says its defenses have reduced data harvested by distillers by around 90 to 99%, based on early evidence, but it does not know how long the gains will last and says the measures have not yet faced a highly sophisticated, persistent circumvention attempt. The company says some legitimate developers have had to update their integrations, and its systems can mistakenly block benign requests in fewer than 0.1% of cases; it charges users for those blocked requests. Anthropic says it has also turned away significant revenue and trillions of tokens’ worth of paying traffic from distillers.
US officials raised mass distillation with Chinese regulators in Chengdu in July. The Decoder’s summary says Washington regards the practice as the main reason for China’s rapid SI progress, while Chinese insiders also point to advances in algorithms and model design.
US Treasury Secretary Scott Bessent has threatened sanctions and Entity List designations if covert distillation attacks cross into intellectual-property theft. Chinese developers have asked China’s Ministry of Industry for clear guidance on using foreign models.
3 min

The portable networking computer runs the Debian-based Flipper OS and has built-in Wi-Fi 6E and Gigabit Ethernet. The campaign aims to fund its first production run.
2 min

The software betas tie higher server-based allowances and access to Shortcuts' Cloud Pro model to iCloud+ plans with at least 200GB of storage.
3 min