[Paek Jongmin's Shockwave] Deepseek on Your PC: Even Chinese Communist Party Issues Answered Smoothly

Open-Source 'Deepseek R1' AI Can Be Freely Installed on PCs
Blocks Personal Data Leaks and Enables AI Utilization
14-Billion Parameter Model Runs on Older NVIDIA RTX 3080 GPU
Provides Answers to Sensitive China-Related Questions When Instal

There is intense interest from users and academia in China’s generative artificial intelligence (AI) ‘R1’ developed by Deepseek, which is said to have been created at a low cost to a level comparable to OpenAI’s ChatGPT in the US. In particular, Deepseek has released R1 as open source, and amid growing concerns over personal information leaks, attempts to install and use R1 on personal computers are increasing.

The popularity of Deepseek, which can be used for free compared to ChatGPT that requires payment, is a global phenomenon. In the US, the Deepseek application (app) ranked number one in downloads, sparking controversy. Concerns spread that various personal data could be transferred to China. As more users flocked, there have been increasing cases where R1 does not provide answers.

However, there is a way to use Deepseek AI at home or in businesses without worrying about security: by installing it directly on a computer.

The advancement of AI technology enabling powerful AI models to run on personal computers marks a significant change. Open-source AI, once led by Meta, Facebook’s parent company, has exploded with Deepseek. Open-source AI models can be downloaded and run, ensuring data privacy while delivering excellent performance. Running AI models in a local environment allows for data security and privacy protection. It enables the use of powerful AI functions without transmitting sensitive information to external servers, allowing companies to safely handle confidential and personal data.

Another major advantage is the ability to continuously use AI functions without relying on an internet connection. Especially on computers equipped with high-performance graphics processing units (GPUs), users can experience having a personal AI supercomputer.

If there is a computer at home for PC gaming, attempts are entirely possible. Running both models smoothly requires appropriate hardware specifications. Deepseek R-1 requires at least 16 gigabytes of RAM and an NVIDIA or AMD GPU. Meta’s LLaMA, for example, can run the 7-billion parameter model with 8 gigabytes of RAM and a basic GPU, but models with 13 billion parameters or more require 16 gigabytes of RAM and a high-performance GPU. Of course, the “the bigger, the better” principle applies. The more powerful the central processing unit (CPU), especially the GPU, the larger the model you can experience.

The reporter succeeded in installing R1 on a computer used by their child. Thanks to the NVIDIA RTX 3080 GPU installed on the reporter’s PC, the R1 challenge was possible. A model with 14 billion parameters was selected to match the GPU’s 10 gigabytes of memory. As the number of parameters increases, R1’s performance improves, but the computer’s performance, especially the GPU memory, must be larger. For a 32-billion parameter model, a GPU with more than 24 gigabytes of memory is required, so the 14-billion parameter model was chosen. First, a tool called Ollama was installed on the PC, then the command ‘ollama pull deepseek-r1:14b’ was executed in the CMD terminal, similar to the old DOS window. After installation, the command ‘ollama run deepseek-r1:14b’ was issued. A message indicating that R1 was running appeared, and questions could be asked.

Meta’s LLaMA also offers models of various sizes and performance levels, allowing users to choose according to their needs and hardware specifications. The installation process is similar to Deepseek’s.

In actual use, R1 stood out for providing detailed reasoning processes for complex questions. Because it shows the reasoning process, it took some time. It was also confirmed that GPU usage soared once R1 started answering.

DeepSeek R1 installed on the reporter's PC is answering questions. When R1 responds, a significant increase in GPU usage can be observed numerically. Photo by Baek Jong-min, Tech Specialist

Since the model used by the reporter was the 14-billion parameter model, the level of answers was somewhat limited. Compared to using the highest-level model via the app, the scope of use seemed inevitably restricted.

Contrary to expectations, it also provided answers critical of the Chinese Communist Party and President Xi Jinping. When asked about the “war crimes of the Chinese Communist Party,” R1 showed a neutral stance but gave substantial answers. The same question could be answered by the US AI Perplexity service using R1. This seemed to reflect the open-source nature. On the other hand, the app version of R1 said it could not answer the same question.

The fact that AI can be used for free on-device without an internet connection is certainly a major change. For example, it can be used as an alternative if ChatGPT service is suspended.

DeepSeek R1 installed on the PC is answering questions about the Tiananmen Square incident and the list of war crimes committed by the Chinese Communist Party. Photo by Baek Jong-min, Tech Specialist

Most difficulties encountered during installation were reportedly resolved through active community support. Professor Jaecheol Kim of the AI Graduate School at KAIST explained, “Using LM Studio, you can use R1 just like using ChatGPT on the web.”

Overseas users also generally gave positive reviews. Especially users who value data privacy highly appreciated the local execution environment.

The AI market is currently developing along two main directions: open-source models and commercial models. In the open-source AI camp, various models have been released, led by Meta’s LLaMA. Besides Deepseek, France’s Mistral and the United Arab Emirates’ Falcon are representative. In the image generation field, Stable Diffusion represents the open-source camp. LG Group has also released Exaone as open source.

Unlike the DeepSeek R1 installed on the PC, the DeepSeek R1 app connected to the internet refused the request to provide the Chinese Communist Party's war crimes list. Photo by Baek Jong-min, Tech Specialist.

The biggest advantage of open-source models is that the code is publicly available, allowing direct inspection of the model’s internal structure and operation. Users can modify and optimize the model according to their needs, and no additional costs occur beyond the initial installation. Especially since they run locally, the risk of data leakage is minimized. It is an open secret that many domestic companies providing AI services already use Meta’s LLaMA. It is a structure that can deliver maximum performance with minimal investment.

In the commercial AI model market, various language models such as OpenAI’s GPT-4, Anthropic’s Claude, and Google’s Gemini are being serviced. In the image generation field, OpenAI’s DALL-E and Midjourney provide services. These commercial models have the advantage of immediate use via web or API without complicated installation processes. They also offer stable performance through continuous updates and optimization, along with professional technical support and documentation.

Local installation of open-source AI models can be an excellent alternative for users who prioritize data security and privacy. Although some technical knowledge is required for initial setup, once overcome, powerful AI functions can be freely utilized. These open-source AI models are expected to continue evolving and be used in increasingly diverse environments. By considering hardware specifications and usage purposes to select the appropriate model, anyone can harness a powerful tool like a personal AI supercomputer. Recognizing this change, NVIDIA has already announced the sale of a compact personal AI supercomputer called ‘Project Digits.’

As competition between the US and China intensifies, consumer benefits may also increase. After the Deepseek shock, OpenAI CEO Sam Altman mentioned, “We are also considering an open-source strategy.” Mark Zuckerberg, CEO of Meta, who has led open-source AI, said on a podcast, “If there is an open-source model that everyone uses, I hope it will be an American model.”