Distillation, a technology that moves high-performance AI models to smaller, cheaper, and easier-to-move models, has become a new issue in the battle for AI supremacy between the United States and China.
Originally, it was a widely used method in research and product development, but now it has spread to the intellectual property and security discussion of "withdrawing capabilities without consent from the cutting-edge models of other companies".
In this article, we will organize the points that companies and developers want to keep in mind about what distillation is and why it has become a US-China problem.
What is model distillation?
Cutting-edge, large-scale AI is called the “frontier model,” and training requires enormous computing resources, data, and investment.
Model distillation is a method of training smaller “student models” using large “teacher model” outputs. The teacher model provides samples of answers, codes, and so on, which are then used as teaching materials for the student model to learn from.
However, the distilled model is not a complete replica of the teacher. It is an image that efficiently learns behavior that works for a specific task, rather than taking over the weight or architecture as it is.
- Lower costs
- Makes inference faster and lighter
- Easy to move even with low-performance hardware
- Easy to make small models that specialize in applications
This makes it an attractive option for companies and governments that want to extend AI to the field, such as devices, factories, automobiles, and closed networks.
Why is “reasoning trace” important?
In recent discussions, there has been a growing interest not only in the final answer, but also in shifting the thought process (reasoning trace) to it.
Florian Trammell, an associate professor at the Technical University of Zurich, likens this to human learning, explaining that it is easier to learn from a collection of questions with a formula than from a final answer alone.
In other words, the view is that the output of AI can include not only "answers", but also the way of solving difficult problems themselves. Access to output is a strategic reason to go beyond just chat usage.
The distillation itself is not illegal
Distillation is a widely used technology in AI training and is not a problem in itself.
There is also public research in the United States that extends small models using the output of advanced models, such as Stanford's Alpaca and Microsoft's Orca. Chinese researchers have also used the output of the US model in their public research.
The big difference is access. Open-weight models allow you to examine and modify parameters. Closed models like ChatGPT and Claude, on the other hand, are designed to only touch through APIs and dedicated UIs.
Why did it become a US-China problem?
The point of contention is not “distillation” per se, but rather the systematic extraction of commercially valuable capabilities from closed models without the permission of the developer.
The US government and major US AI companies have accused Chinese companies of using distillation to extract capacity from US models. Regulatory focus is also expanding from advanced semiconductor and export regulations to model learning methods and intellectual property.
One of the triggers for heightened vigilance has been reports that the Kimi K3 Moonshot AI (Dark Side of the Moon) has been perceived as being dulled by cutting-edge US models.
- Anthropic accuses DeepSeek, Moonshot, MiniMax and others of large-scale activities to acquire capabilities from Claude
- OpenAI also stated that it had detected an attempt to use its model for distillation purposes.
- US Treasury Secretary Says Possible Sanctions If Tech Piracy Is Confirmed
On the other hand, the Chinese side has argued that the US criticism is unreasonable interference, and some reports have rebutted that "US companies are also distilling the Chinese model". There is an overlap between technology and geopolitical frameworks.
Why is it difficult to regulate
Chip export regulations are easy to target hardware and production lines, while distillation can occur with access to output and repeated experiments, making it difficult to draw lines.
The policy side's concern is that it is difficult to define "how far is legitimate research and product improvement, and where is organizational capability extraction?"
There is also a concern that if useful behaviors can be absorbed from the output alone, the diffusion of capacity cannot be completely contained by tip limits or by not disclosing model weights alone.
What companies and developers need to think about
It is not another human resource for Japanese companies or development teams. The more you use the output of the Frontier API to develop internal and specialized models, the more important the terms of use, monitoring, and accountability.
- Check if downstream model training and distillation is allowed in the API Terms of Service
- Avoid usage that looks like a lot of power collection or scraping
- Record the learning path and data origin of the internal model
- Design a balance between closed API dependencies and open models/in-house data
- Clarify distillation and performance, safety and governance trade-offs to reduce costs
Particularly in business automation and AI agents, it is tempting to turn to small, inexpensive models. Before that, you need to decide which outputs to use, for what purpose, and how far you can use them for learning.
Practice of AI introduction and model selection
The distillation controversy is more about access control and procurement than competition for model names. The following decisions will take effect in the implementation of the company.
- Continue to use the strongest hosted APIs
- Is it compressed into a small and specialized model for narrow work?
- Reduce vendor dependency around open models
- How to divide the privilege log cost at the connection point with the existing system
Makoto Tejima supports the implementation of the latest AI models, including design based on the terms of use, integration with existing systems, authority, monitoring, and cost design.
Please feel free to contact us if you have any questions about how to proceed, such as internal AI, agent implementation, or adding AI to existing web systems.
Recap
Distillation of AI models is a method of efficiency that has been used for many years. What is at stake now is a vigilance against the systematic extraction of capabilities without consent from closed, cutting-edge models, rather than the technology itself.
Sparked by Kimi K3 and others, US-China disputes have spread from chip regulation to the discussion of learning methods, intellectual property, and sanctions. The company side needs to see the API conditions, data origin, and model selection in a set.
Please feel free to contact us if you have any questions about system development using AI, deployment design that balances cost and governance, quotes, and how to proceed.
