Frontier Models
PLA Researchers Distill OpenAI and Anthropic Models for Military Drone AI
Institutions tied to China's military have applied model distillation to outputs from US frontier models to create compact domestic AI for drones, navigation and maritime operations while avoiding hardware restrictions.
Model distillation is the technique of transferring reasoning and outputs from large frontier AI systems into smaller domestic models for specialized local deployment.
Chinese military researchers have used outputs from leading U.S. artificial intelligence models developed by OpenAI and Anthropic to train domestic AI systems to advance China's defence capabilities. This method permits creation of compact models that operate without advanced foreign chips. The approach relies on software layer access rather than hardware imports. Reuters examined dozens of papers compiled with assistance from the Jamestown Foundation. The review covered publications from 2024 through 2026.
What scale of activity appears in the reviewed Chinese academic papers and patents?
A Reuters review identified more than 80 Chinese academic papers and patents showing widespread use of model distillation by PLA-linked institutions. The papers originate from multiple military research centers. Many authors hold affiliations with the People's Liberation Army. The documents detail extraction of knowledge from commercial US models. These efforts span several years of published work.
The Jamestown Foundation supplied additional research that supported the Reuters analysis. Reviewers found consistent patterns across the documents. Researchers first query the US models. They then convert responses into training material for smaller Chinese systems. The resulting models run on domestic infrastructure. This sequence appears repeatedly in the collected papers.
How did PLA Unit 96941 researchers apply GPT-3.5 outputs in their published work?
One paper published last year by researchers in PLA Unit 96941 described using OpenAI's GPT-3.5 to summarize sensitive military source code. The team generated summaries from the model. They then trained a domestic model on those summaries. The final system operates entirely within Chinese military networks. No external connectivity is required after training.
The method allows retention of useful reasoning patterns without direct model access. Unit 96941 focuses on specialized defense tasks. The distilled model supports internal code analysis functions. Publication details confirm the workflow. The approach reduces dependence on restricted hardware components.
What application did North University of China pursue with Claude 3 Haiku?
Researchers at the North University of China used Anthropic's Claude 3 Haiku to generate synthetic training data for a text classification model for social media monitoring and content moderation. The synthetic data replaced or supplemented real world examples. The resulting classifier performs content analysis tasks. The work supports monitoring functions within military contexts. The paper outlines the data generation step in detail.
Claude 3 Haiku served as the source for labeled examples. The smaller model learns from the generated set. Deployment occurs on local servers. The technique demonstrates transfer of classification capabilities. North University of China maintains ties to defense research programs.
What technical approach did the 2024 paper from the National University of Defense Technology describe?
A 2024 paper from the PLA's National University of Defense Technology described using distillation to shrink an image-processing model for deployment on unmanned aerial vehicles for real-time analysis and navigation in communications-denied environments. The original model was compressed through distillation. The smaller version retains core image analysis functions. It operates on UAV hardware without constant external links. Navigation and analysis occur in real time.
The distilled model processes visual data onboard. Performance targets include denied communications scenarios. National University of Defense Technology conducts extensive UAV research. The paper provides implementation specifics for the compression step. The result supports autonomous operations in contested areas.
How was distillation used by the Academy of Military Sciences in maritime simulations?
Researchers at China's Academy of Military Sciences used distillation to run a target-recognition model on tactical hardware during simulated maritime operations involving drones, ships and unmanned submarines. The model identifies targets in the simulated environment. It functions on limited hardware platforms. The simulation includes multiple unmanned systems. Distillation enabled the required performance level.
Academy of Military Sciences focuses on operational integration studies. The distilled model supports coordinated maritime scenarios. Tactical hardware constraints guided the compression choices. The paper reports results from the simulated exercises. This work aligns with broader PLA efforts in unmanned systems.
What market and stakeholder implications follow from these documented practices?
The documented distillation activities affect US export control policies on advanced chips. Chinese entities obtain reasoning capabilities without direct hardware access. Domestic production of compact models increases. Stakeholders in the US AI sector face questions about output monitoring. Policy responses may target software layer safeguards.
Military applications gain from the reduced hardware requirements. Drone and cyber operations benefit from the compact systems. International technology transfer discussions may intensify. Companies must evaluate access policies for their models. The pattern suggests continued interest in distillation techniques.
What expert reactions address the transfer of reasoning capabilities?
Teaching a model the right answer is one thing but teaching it the reasoning behind the answer is much harder. These papers show Chinese military-linked researchers are trying to transfer that expensive, proprietary reasoning from Western models into smaller systems they can control and deploy locally.Sunny Cheung, Jamestown Foundation fellow
Anthropic said it does not provide commercial access to Claude in China or to Beijing-controlled firms and uses monitoring systems to detect policy violations. The company added that distilled models may lose the original systems' safety safeguards, potentially allowing sensitive capabilities to be transferred to models beyond its control. The statement highlights risks associated with distillation outputs.
What developments are expected next in distillation research and related policy?
Further papers from PLA-linked institutions are likely to appear in coming years. Refinements to distillation methods will continue. Policy discussions around model access controls may expand. Companies may strengthen detection of distillation attempts. Military applications of compact models will receive additional attention.
The Jamestown Foundation continues to track related publications. Reuters maintains coverage of the topic. Additional examples of distillation for defense uses may surface. Stakeholder responses will shape future access rules. The field of military AI distillation remains active.
| Institution | US Model Used | Military Application | Publication Details |
|---|---|---|---|
| PLA Unit 96941 | GPT-3.5 | Summarizing sensitive military source code for local training | Paper published last year |
| North University of China | Claude 3 Haiku | Synthetic training data for social media text classification | Recent paper on monitoring models |
| National University of Defense Technology | Distilled image-processing model | UAV real-time analysis and navigation in denied environments | 2024 paper |
| Academy of Military Sciences | Distilled target-recognition model | Maritime operations with drones, ships and unmanned submarines | Simulation study on tactical hardware |
- Query the frontier model through available software interfaces to obtain outputs or summaries.
- Convert the obtained responses into training examples suitable for a smaller model.
- Train the compact domestic model on the extracted knowledge and reasoning patterns.
- Deploy the resulting system on local military hardware for drone navigation or target recognition tasks.
- Verify operation in simulated or real environments without reliance on restricted chips.
Frequently asked
How does model distillation allow bypassing of chip export controls?
Model distillation transfers reasoning from US models into smaller systems that run on domestic hardware. Researchers avoid the need for advanced chips by using software outputs for training.
Sources
- Reuters — Chinese military researchers have used outputs from leading U.S. artificial intelligence models developed by OpenAI and Anthropic to train domestic AI systems to advance China's defence capabilities.
- Jamestown Foundation — A review of dozens of Chinese academic and industry papers published between 2024–2026 provides evidence of PLA-linked institutions distilling OpenAI’s GPT-3.5 to create smaller models.
- @n7ekg — PLA-linked institutions use model distillation to build compact AI systems for drones and autonomous navigation, bypassing US chip export controls by leveraging software-layer access.