AI in Defense: Analyzing China’s Parade Weapons

On 3 September 2025, China held the Victory Day Parade and revealed some of its advanced weaponry systems.

As someone who grew up with Sci-Fi comics and movies, I am interested in “laser weapons” and “missiles” that can shoot down satellites in orbit.

Also, as a big fan of Beloved Oppressor cum Supreme Leader Admiral-General Aladeen of the Republic of Wadiya (a fictional character from a 2012 film “The Dictator”), I am interested in China’s “pointy weapons” because “pointy is scary”, as Aladeen once said.

So, I tried to find a list of pointy scary weapons China showed off during the parade, but I could not find a comprehensive list and/or organized information. This made me decide to take matters into my own hands.

When I browsed YouTube, I found a video titled “All 76 weapons at China 2025 military parade explained. 47 are brand new.”, made by Binkov’s Battlegrounds channel. Also, I found CNA’s clip titled “China unveils latest missiles, weapons at 2025 Victory Day Parade” interesting.

The video from Binkov’s Battlegrounds channel is, by far, the most comprehensive, as much as I know.

Please note that I am not a weapons expert, and do not have in-depth knowledge of defense weapon systems. I am a mere researcher from a Southeast Asian country who loves structured data/information and working with machines to make something useful and fun.

The following are the steps I’ve taken for building a data analysis pipeline for defense and security analysis work. The processes: Gather OSINT > Process Data > Generate Information > Create Output for Further Analysis and Triangulation by Domain Expert.

My workflow is simple. I hope defense and security analysts and researchers who want to make use of AI and contemporary technology, as well as hobbyists, will find this simple workflow useful. For those who are deeply doing serious research in this domain, using cutting-edge technology, I understand my workflow is just an entry-level one.

The following are the steps I have taken.

Using a Python package called “yt-dlp”, a feature-rich command-line audio/video downloader, I downloaded the video from Binkov’s Battlegrounds channel.

Then, I transformed the video into images, using OpenCV, the world’s biggest computer vision library. As I captured each frame of the video, I received over 36,057 frames/images. This is a rich image dataset derived from OSINT.

Afterward, using OpenCV and Python’s built-in packages, I grouped similar images, and put grouped images into multiple folders/directories. As I ran the tasks using a computer with 16GB of RAM, this process took several minutes, and my PC was almost frozen while processing. At the end, 36,057 images were grouped into 2,384 folders (i.e. 2,384 categories/classes/groups).

The grouped images are quite useful for both manual analysis, as well as AI-assisted analysis processing. For instance, by exploring group folders, I was able to find the interesting groups of images (interesting means “pointy weapons” in my case).

Then, I used the OpenAI “gpt-4o” model to analyze the images (what is visible, identify model/type, system profile). The role assigned for the analysis task is “a defense and security expert with deep knowledge of modern defense systems and weaponry”.

The image below is the pointy weapon I am interested in, and a concise profile report generated by OpenAI (gpt-4o model).

# Output in Markdown 

(1) What is visible
- Multiple military vehicles equipped with missile launchers.
- Each vehicle carries two missiles.
- Visible markings on the missiles read "HQ-16C."
- The vehicles display military insignia and numerical identifiers (e.g., HZ122, HZ123).

(2) Identify model/type
- The system displayed is the HQ-16C, a Chinese medium-range surface-to-air missile (SAM) system.
- The vehicles are likely transporter erector launchers (TELs) specifically designed for the HQ-16 series.

(3) System profile
- **Role**: The HQ-16C is a medium-range air defense system designed to target aircraft, helicopters, and other airborne threats.
- **Capabilities**: It provides all-weather, day-and-night air defense capabilities with a range of approximately 70 kilometers and an altitude of up to 18 kilometers. The system can engage multiple targets simultaneously, enhancing battlefield air defense coverage.
- **Notable Features**: The HQ-16C is known for its vertical launch system, which allows for 360-degree coverage. It incorporates advanced radar and guidance systems for improved target tracking and engagement accuracy.
- **Service Status**: Entered service around 2019, the HQ-16C is widely deployed within the People's Liberation Army (PLA) as part of China's modernized air defense network. The exact inventory is unknown but is speculated to be in the thousands, reflecting its significant role in China's strategic air defense strategy.

As I mentioned before, being a person who does not have much knowledge in weapon systems, the information generated by OpenAI is quite useful for me. It helps me a lot to understand the lethality of “pointy scary weapons”.

Yet, it is important to note that the AI-generated output may not be perfect. If human domain experts analyze the outputs, accuracy will be improved.

The grouped images can be batch processed using OpenAI, and the outputs can be merged to generate a synthesis report. For this exercise, I only processed a few images, and stopped generating reports for all groups to save my OpenAI API credits.

Individual researchers and/or institutes interested in doing AI-assisted analysis to understand the weapons exhibited at China’s V Day parade can use the same approach. Processing costs (API fees) will be less than US\$ 100-200 for running all grouped images in batch. The workaround to save API costs is manually selecting interesting images (perhaps 100-200 selected images).

The lessons learned from this exercise are as follows.

The use of AI computer vision saves time, reduces cost, reduces human error, and improves efficiency.
AI can help human researchers/analysts in a wide range of activities in the research pipeline.
AI can do low-value tasks (such as classification/grouping images, analyzing similarities/differences, etc.) so that human researchers/analysts focus on high-value tasks (strategic thinking, formulating worst case scenarios, etc.)

The image dataset I generated is 12.5 GB, and let me know if you want the image dataset. Please reach out to me if you have any questions or need clarification. Any suggestions and comments are welcome.