An Analysis of China's First Military LLM
China: Technosphere provides a biweekly deep-dive into China’s technological control at home and influence abroad.
Any translations into English are unofficial and informal. All Chinese-language links in this brief are safe to click on. They will lead you to search enginge results or an archived version of websites. Do not visit the actual websites, as this might expose you to being tracked by the Chinese state.
Subscribe for free, or support China: Technosphere with a paid subscription to unlock the full archive of previous posts and the database of featured companies.
Biweekly Feature: PLA Reasoning
Key takeaway: In 2023, China’s first military LLM was launched. It is used for planning, force allocation, intelligence analysis, and operational optimization, trained on sources such as open-source, signals, and human intelligence. It supports scalable model sizes (requiring specialized Huawei hardware) and is already used by PLA units to accelerate decision-making processes to seconds. The emergence of artificial decision-making is likely to have a sizeable impact on deterrence. Influencing how AI systems think, may become as strategically important as influencing human leaders.
In 2023, DataExa launched the Tianji Military LLM. It has 5 core functionalities:
Military Q and As
Combat Mission Planning
Force Allocation
Military Operations Optimization
Intelligence Compilation
The company draws on a variety of sources, including open source intelligence, signals intelligence and explicitly states that human intelligence (HUMINT) is part of the information sources.
Depending on how much power is needed the company offers models with the following parameter sizes: 7billion, 14B, 72B, or 130B+. It allows for the concurrent deployment of mulitple LLMs and DeepSeek integration. To run the 130B LLM locally users need what appear to be 6 chips from the Huawei 910 Ascend series with 64G of High Bandwidth Memory.
The LLM is used by a PLA theater command to optimize and automate ammunition transportation and by a PLA defense academy to plan time-sensitive target missions.1 A PLA command academy uses it to answer questions like what is “the mobility performance of a certain type of tank in high-altitude environments and its correlation with historical battle examples.”
Here some more questions that are highlighted in the Tianji LLM.
What radars are equipped on the USS George Washington aircraft carrier?
What types of aircraft are on the Nimitz-class nuclear-powered aircraft carriers?
What weapons and equipment does the Kunming guided-missile destroyer carry?
A PLA joint operations command center used the Tianji LLM to reduce the threat assessment-target allocation-attack damage assessment chain to 8 seconds.
In a way DataExa aims to mirror US approaches to deploying AI in the military.
DataExa appears to provide an amalgamation of capabilities provided by Palantir (structuring, labeling, and preparation of defense data for AI training), OpenAI’s “Codex” (multi-agent orchestration platform), and Anduril’s “Lattice” (making tactical data ready for AI training) capabilities. DataExa’s AI agents create advanced military simulation environments for game-playing.
Figure 1: The DataExa all in one mlilitary LLM machine (Source: DataExa)
Figure 2: A U.S. Military Facilities Encyclopedia (Source: DataExa)
Figure 3: DataExa’s Digital Battlefield Map (Source: DataExa)
When AIs analyze each other
DataExa’s system devises strategies for the PLA based on China’s own military power indicators and those of its adversaries, historical battlefield lessons, strategies and tactics of its adversaries. For democratic countries, it may be just as important to understand how systems like those of DataExa think as it is to understand how PLA leaders think, or how DataExa believes democracies’ military AI systems think. Subtly influencing these systems will play a central role in 21st-century geopolitics.
What if the Tianji Military LLM could be led to conclude that invading a neighboring country does not make sense? What if deterrence were as much about influencing the reasoning of PLA LLMs as it is about influencing Xi Jinping?
Company spotlight: DataExa(厦门渊亭信息科技有限公司)
DataExa is located at 8th Floor, No. 61-1 Wanghai Road, Software Park II, Siming District, Xiamen. Hong Wanfu (洪万福) is the founder and CEO of DataExa. He established the company in 2014. DataExa is not only active in the military industry but also in supporting financial (People’s Bank of China) and telecommunications industries (China Mobile) with its software.
Figure 4: Hong Wanfu (Source: Xiamen Evening News)
In March 2026, DataExa launched MilClaw, a military grade version of OpenClaw, an autonomous AI agent.
Figure 4: MilClaw architecture (Source: DataExa)
The following entities are DataExa customers: People’s Liberation Army Army, Science and Technology Commission of the Central Military Commission. See footnote 2 for a full list of government departments, military academies, and university customers.2
This procurement document shows DataExa supplying a “Visual Knowledge Weaving and Application Training Platform Based on [a] Large Model” to what is a customer in Changsha, Hunan province. In this case, the customer might be the Changsha-based National University of Defense Technology.
Thank you for reading China:Technospehere.
The information presented in this biweekly brief is intended solely for informational and academic purposes. It reflects the author’s independent research, analysis, and opinions, and does not constitute legal, financial, investment, or professional advice.
While every effort has been made to ensure the accuracy and timeliness of the content, no guarantee is provided regarding the completeness or correctness of the information. Readers are encouraged to conduct their own due diligence before drawing conclusions or making decisions based on this publication. The author disclaims any liability for actions taken or not taken based on the contents of this brief.
Time-sensitive-targets. Those targets requir[e] immediate response because they pose (or will soon pose) a danger to … forces or are highly lucrative, fleeting targets of opportunity whose destruction is of high priority to achieve campaign objectives.
People's Liberation Army Army, Science and Technology Commission of the Central Military Commission, People's Liberation Army Navy, Joint Logistics Support Force of the Central Military Commission, Chinese People's Liberation Army Air Force, National Defense University of the Chinese People's Liberation Army, Chinese People's Liberation Army Rocket Force, Academy of Military Sciences of the Chinese People's Liberation Army, Strategic Support Force of the Chinese People's Liberation Army, National University of Defense Technology of the Chinese People's Liberation Army, Equipment Development Department of the Central Military Commission, PLA Army Command Academy, Training and Management Department of the Central Military Commission, Air Force Command Academy of the Chinese People's Liberation Army, Joint Staff Department of the Central Military Commission, PLA Army Armored Force Academy.