Model distillation - a technique that allows developers to train compact, lower-cost AI systems using the outputs of much larger models - has moved from routine research practice to a geopolitical flashpoint in the intensifying competition between the United States and China over artificial intelligence.
At its core, distillation relies on a two-step relationship. A large, resource-intensive "teacher" model generates responses, examples or solutions. Those outputs then serve as training material for a smaller "student" model. The student is not a direct copy of the teacher; it does not receive the teacher's weights, architecture or full set of capabilities. Instead, it internalizes selected behaviors that make it able to complete particular tasks more efficiently.
The largest AI systems - often called frontier models - demand substantial compute, data and investment to build and operate. Distillation offers a pathway to take some of the frontier model's task performance and compress it into systems that can run on less powerful hardware. That cost and efficiency advantage makes distilled models attractive for deployment outside massive data centers - for use in devices, factories, vehicles and private networks - and for organizations that lack the resources to run or license frontier models directly.
Interest in distillation has broadened as researchers aim to transfer not only final answers but also the process by which those answers are reached. So-called "reasoning traces" - a model's intermediate steps or chain-of-thought in arriving at a solution - can act as a form of instruction. Florian Tram e8r, an assistant professor at ETH Zurich who studies machine-learning security, compared the approach to learning mathematics: receiving final answers alone is far less instructive than getting detailed step-by-step solutions, which show how to approach and solve problems.
As reasoning traces gain prominence, access to model outputs becomes more sensitive because they can reveal methods that sophisticated systems use to handle complex tasks. That sensitivity is a core element of the current dispute.
Distillation itself is a long-standing tool in AI research and development. U.S. projects have employed it publicly, including Stanford University's Alpaca work and Microsoft's Orca research, both of which used outputs from larger models to improve smaller ones. Researchers in China have likewise used outputs from U.S. models in public research efforts, such as projects aimed at creating Chinese-language instruction models.
What distinguishes different uses of distillation is the degree of access to the underlying model. Open-weight models allow researchers to inspect and modify parameters directly. By contrast, closed models - for example certain commercial systems accessed only through proprietary interfaces or APIs - remain under company control. Companies that operate closed models typically restrict how their systems are queried and how outputs may be harvested.
The current controversy is not about distillation as a technique but about the scale and authorization of extracting outputs from closed, proprietary models. Several U.S. AI companies say they have identified systematic efforts to collect outputs from their closed systems and use those outputs to derive capabilities for competitive products.
Anthropic has levelled accusations at Chinese actors including DeepSeek, Moonshot and MiniMax, alleging large-scale campaigns to obtain capabilities from Anthropic's Claude models. According to the company, those campaigns targeted skills such as advanced reasoning and software engineering. OpenAI has also reported detecting attempts by Chinese actors to use its models in ways linked to distillation.
No Chinese firms have publicly accused U.S. rivals of conducting similar distillation campaigns against closed-source models.
What the dispute centers on
- Whether outputs from closed, proprietary models can be systematically harvested without consent to reproduce valuable behaviors.
- How the transfer of reasoning traces - not just final answers - might enable smaller models to replicate complex problem-solving methods.
- Differences in access models: open-weight systems permit inspection, while closed systems restrict direct examination and control distribution via interfaces or APIs.
The debate has practical implications for a range of sectors. Technology companies that build and sell AI models, cloud service providers, and industries seeking to embed AI in devices, factories and vehicles are directly affected. Governments and private network operators also face choices about how to acquire and protect advanced capabilities.
While distillation can lower hardware and operational costs and broaden the reach of AI, the contentious context around unauthorized large-scale extraction has elevated it from a technical method to a matter of commercial control and national competition.
What remains limited in public discussion
Public accounts to date focus on the distinction companies draw between legitimate research and what they view as systematic capability harvesting from proprietary systems. The available information describes accusations by some U.S. firms and the identification of campaigns targeting specific capabilities, but it does not include formal cross-accusations from Chinese companies about similar activity.
Given the evolving nature of AI research and deployment, the discussion about distillation highlights tensions between open scientific practice and the protection of commercially developed, closed systems. The balance between those priorities continues to shape how companies and governments approach access to model outputs and the reuse of those outputs for training smaller systems.