17 Interesting Facts About DeepSeek, the Latest Chinese AI Competing with ChatGPT

DeepSeek has recently become a topic of discussion due to its ability to challenge the dominance of AI technology from the US. This AI developed by a Chinese company shares many similarities with ChatGPT from OpenAI, Gemini from Google, and Claude from Anthropic. DeepSeek is also capable of understanding and responding to various user commands in a format similar to other AI chatbots.

What sets DeepSeek apart from other AI models is its efficiency in development. With its latest model, DeepSeek has the potential to surpass ChatGPT in intelligence while being much cheaper to develop, making it a challenge to the dominance of US-based AI. Here are some interesting facts about DeepSeek from China that you should know:

1. Still Young

As an AI model developer, DeepSeek is still quite young, only about two years old. This AI startup is based in Hangzhou, Zhejiang, China, and was founded in 2023.

2. Founded by Liang Wenfeng

DeepSeek emerged from an initiative by High Flyer, a hedge fund from China, and is led by Liang Wenfeng. Wenfeng, born in 1985 in Zhanjiang, Guangdong Province, China, is a graduate in Electronic Information Engineering from Zhejiang University.

Although the AI model he developed is still relatively new, Wenfeng is optimistic that DeepSeek can push China's progress in AI technology innovation. Like other AI developers, he has a long-term vision to create an AI model that can achieve Artificial General Intelligence (AGI), an AI with human-like intelligence.

3. Outperformed ChatGPT in the App Store

DeepSeek is now available in various formats, including as a mobile app. On Monday (January 26, 2025), the DeepSeek mobile app managed to top the free apps category in the App Store in 111 countries. Its rise in popularity was significant, as just a few days earlier, DeepSeek was ranked 31st.

On the App Store, DeepSeek even outperformed other similar apps like ChatGPT. According to data from Appfigures, DeepSeek also topped the free app list in Google Play Store in 18 countries.

4. Has Two AI Models

Since its establishment in 2023, DeepSeek has developed several AI models, all named "DeepSeek." The two latest models they released are DeepSeek V3 and DeepSeek R-1.

DeepSeek V3, launched in December 2024, is a Mixture-of-Experts (MoE)-based model with a total of 671 billion parameters, but only 37 billion parameters are activated per token during inference, making it highly efficient. 

This model can handle context windows up to 128,000 tokens and produce outputs up to 8,000 tokens, suitable for tasks like answering everyday questions and creative content generation.

Meanwhile, DeepSeek R-1, released on January 20, 2025, has gained attention for its superior intelligence. Developed based on DeepSeek V3, this model features enhanced reasoning capabilities through reinforcement learning. 

DeepSeek R-1 is able to demonstrate its thought process before reaching conclusions and solve complex problems better. It also has a larger output capacity compared to DeepSeek V3, with up to 32,000 tokens.

5. Born Amid US AI Chip Export Restrictions

DeepSeek emerged in the midst of the US-imposed AI chip export restrictions. Currently, the AI industry is still dominated by the US, with most AI chips being produced by major companies like Nvidia.

In mid-January, the US government set new regulations tightening oversight of AI chip exports from industry giants like Nvidia and AMD to global markets.

This policy aims to control the distribution of advanced AI technology, especially to countries outside of US allies and partners, and maintain the US's dominance in the global AI race.

Additionally, this regulation simplifies the export licensing process, closes loopholes in smuggling, and implements new security standards to prevent such technology from falling into unwanted hands.

The new rules also limit AI chip exports to countries considered potential threats to US national security, such as China, Russia, Iran, and North Korea. As a result, DeepSeek has faced difficulties obtaining the latest, cutting-edge AI chips.

6. 10 Times Cheaper Than ChatGPT

The development of DeepSeek has been more efficient in terms of time and cost compared to GPT-4, the popular AI model used in OpenAI's ChatGPT.

DeepSeek R-1 required about two months of training at a cost of around $6 million USD (about Rp 97 billion), which is 10 times cheaper than ChatGPT.

By comparison, the development of GPT-4 cost about $63 million USD (around Rp 1 trillion) and took several months to a year.

7. Only Uses Nvidia H800 Chips

The efficiency of DeepSeek’s development is due in part to the use of more affordable Nvidia H800 GPUs for training. In contrast, GPT-4 was developed with the higher-end and more expensive Nvidia H100 chipset.

Chinese AI models like DeepSeek cannot access the latest Nvidia H100 chips due to the US export restrictions, especially toward countries considered at risk, such as China.

Besides using cheaper chips, the number of GPUs used for training DeepSeek is also more efficient. DeepSeek only utilizes 2,048 Nvidia H800 GPUs, whereas GPT-4's training could involve tens of thousands of Nvidia H100 GPUs.

8. More Efficient and Optimal in Resource Usage

DeepSeek employs innovative technology approaches to enhance the efficiency and performance of their AI models. In its development, DeepSeek uses Mixture-of-Experts (MoE) and Chain of Thought (CoT) techniques.

MoE is an architecture that allows large models like DeepSeek V3, which has 671 billion parameters, to activate only 37 billion parameters when processing each token. This method allows for more resource-efficient usage without compromising performance.

Additionally, DeepSeek R-1 was trained using the Chain of Thought (CoT) technique, which breaks down complex questions into smaller steps before providing a final answer. This approach not only improves the logic and accuracy of responses but also helps the model identify and correct logical errors or data hallucinations during its thinking process.

9. Excels in Various Benchmark Platforms

Despite being developed with more efficient costs, DeepSeek's capabilities are still impressive. This Chinese AI model is claimed to outperform several other leading AI models, such as Claude from Anthropic, Llama from Meta, and GPT from OpenAI, in various benchmarks.

For example, in the context understanding test (DROP, 3-shot F1), DeepSeek V3 scored 91.6 points, higher than Llama 3.1 (88.7), Claude 3.5 (88.3), and GPT-4 (83.7).

Additionally, in solving international-level math problems, such as AIME 2024, MATH-500, and CNMO 2024, DeepSeek V3 recorded scores of 39.2, 90.2, and 43.2 points, respectively.

By comparison, Llama 3.1 scored 23.3, 73.8, and 6.8 points; Claude 3.5 scored 16.0, 78.3, and 13.1 points; while GPT-4 scored 9.3, 74.6, and 10.8 points.

DeepSeek also claims that DeepSeek R-1 can compete with and even surpass OpenAI's latest AI model, OpenAI O1, in various benchmarks, including context understanding and problem-solving.

10. Relying on Local Talent

DeepSeek's development is entirely carried out by local talent. Wenfeng emphasizes that AI innovation doesn't need to depend on experts from outside of China. All DeepSeek employees are from within China, with the majority being recent graduates from prestigious Chinese universities and experienced young talents in the AI field.

Wenfeng also acknowledges that there may not be many local AI experts currently on par with global talent. However, he is committed to nurturing and developing their skills to compete internationally.

11. DeepSeek Made Open Source

Despite having capabilities that can compete with other popular AI models, DeepSeek embraces an open approach. DeepSeek R-1 is developed as an open-source project, unlike ChatGPT, which is closed-source. With this development model, the code of DeepSeek R-1 is accessible, usable, and modifiable by anyone.

This open approach allows the global developer community to contribute to refining and advancing DeepSeek, thus accelerating its growth and evolution.

12. Causing U.S. Tech Stock Market Drop

The arrival of DeepSeek signals to the U.S. that its dominance in the AI industry might not last long. The launch of DeepSeek's latest AI model on January 20th caused significant impacts on the stock market, particularly for U.S. tech companies.

On Monday (January 27, 2025), shares of U.S. tech companies saw a sharp decline. Nvidia's (NVDA) stock, a major supplier of AI chips, plummeted nearly 17 percent, resulting in a market value loss of $588.8 billion.

This became the record for the highest market value loss in a single day, surpassing the previous record set by Meta almost three years ago with a drop of $240 billion.

Not only Nvidia, but other tech companies like Meta (META) and Alphabet (GOOGL) also saw significant stock price declines. Shares of Oracle (ORCL), Vertiv, Constellation, NuScale, and other data center companies also dropped sharply.

13. Causing Market Doubts About U.S. Tech Capabilities

Keith Lerner, an investment analyst at Truist Financial, stated that DeepSeek's arrival has made the market doubt the capabilities of U.S. companies in the AI tech industry, which were previously considered superior.

Lerner mentioned that the launch of DeepSeek's model made investors question the edge U.S. companies held, wondering how much they were spending and whether their expenses would generate profit or just be wasteful.

On the other hand, Charu Chanana, head of investment strategy at Saxo, argued that with more efficient development, DeepSeek’s growing potential might attract investor interest, as it appears capable of providing promising growth.

14. Leading Crypto Investors to Sell Assets Massively

In addition to affecting the stock market, DeepSeek's presence also impacted the cryptocurrency market. On Monday (January 27, 2025), Bitcoin's price fell by 7 percent, dipping below $98,000 (approximately Rp 1.58 billion) per coin.

Crypto analytics platform Coinglass reported that on that day, investors massively sold their crypto assets, with total liquidations reaching $861.48 million (around Rp 13.9 trillion).

Crypto experts indicated that this massive sell-off was related to DeepSeek's presence, which caused overvalued tech stocks to undergo a revaluation.

15. Raising U.S. Awareness

The arrival of DeepSeek from China has made the U.S. government wary of its potential impact. White House Press Secretary Karoline Leavitt confirmed that U.S. officials are evaluating the potential threat posed by this AI model.

Concerns have arisen regarding the distillation techniques used, triggering discussions about the need for tighter measures to prevent misuse of U.S. AI technology. U.S. tech companies are also becoming more cautious of DeepSeek, with Microsoft CEO Satya Nadella warning AI companies to be careful.

Meta, which is also developing the AI model Llama, is reportedly planning an analysis of DeepSeek’s technology. Meta is interested in investigating how DeepSeek could cut development costs and the data used in its development process.

16. Welcomed by Trump and ChatGPT's CEO

Despite concerns from some parties in the U.S., several important figures have positively welcomed the arrival of DeepSeek from China. Former President Donald Trump and ChatGPT CEO Sam Altman are among those who have praised DeepSeek’s presence.

Trump stated that DeepSeek could serve as a warning for U.S. companies to enhance their competitiveness. According to him, China’s smarter and more cost-efficient AI is a positive development.

With a similar view, Altman also saw DeepSeek as a positive push in the AI competition. Its arrival encourages OpenAI to develop better models.

17. Launching the Janus Pro AI Image Generator

Amid the popularity of DeepSeek R-1, DeepSeek has just launched a new AI model capable of generating images multimodally, both from text and image prompts, known as the AI Image Generator. Previously, DeepSeek R-1 and DeepSeek V-3 lacked the ability to generate images from text and images.

DeepSeek's new AI Image Generator model is named Janus Pro. Janus Pro is claimed to outperform similar AI models like DALL-E 3 from OpenAI (the parent company of ChatGPT) and Stable Diffusion. Currently, Janus Pro is available for download via the AI platform Hugging Face.

That’s 17 interesting facts about DeepSeek, the latest AI from China that is challenging ChatGPT. With its efficient technology and lower development costs, DeepSeek has attracted global attention and poses a serious challenge to the dominance of AI from the United States. Although it is a newcomer, DeepSeek has the potential to reshape the AI industry with its open-source approach and innovations that could significantly accelerate the evolution of AI technology.

Also, check out @felixindo on Instagram for more insights about DeepSeek and the latest developments in AI technology.