What is DeepSeek and why everyone is talking about it

A new AI development company is revolutionizing the digital landscape. Its name is DeepSeek, and it is a Chinese startup founded in 2023 specializing in the creation of open-source models. It has been the most recent prototypes, DeepSeek R1 and DeepSeek V3, which have brought the Chinese developer into the spotlight, positioning it as a strong competitor against other industry giants.

The fact that this open-source AI is as competitive as OpenAI’s technology and requires less powerful chips has led to questions about the technological leadership of U.S. companies such as Nvidia, which saw its shares fall by 17% and lost approximately 660 billion dollars. This caused Nvidia to lose its title as the world’s most valuable company, being surpassed by Apple (1st) and Microsoft (2nd). Other tech giants were also affected, with Alphabet falling by 4% and Microsoft by 2%.

Furthermore, there was also a significant downturn in stock markets, especially in the Nasdaq, which plummeted by 3.77%. On Monday, the 27th, the futures of the Dow Jones declined by 1.04%, while those of the S&P 500 fell by 2.29%. These negative reactions began to manifest last Friday, when U.S. stocks fell in the market.

Currently, there are a series of restrictions imposed by the United States, prohibiting the sale of advanced chips to Chinese companies. These have not prevented DeepSeek from advancing in its technological research. Journalist Holger Zschaepitz explained on his social networks that the fact that DeepSeek seems to «have built an innovative artificial intelligence model at an extremely low cost and without access to state-of-the-art chips, calls into question the utility of the hundreds of billions of dollars in capital expenditures being invested in this industry».

China’s #DeepSeek could represent the biggest threat to US equity markets as the company seems to have built a groundbreaking AI model at an extremely low price and without having access to cutting-edge chips, calling into question the utility of the hundreds of billions worth of… pic.twitter.com/wMVyczpRgT

— Holger Zschaepitz (@Schuldensuehner) January 24, 2025

What is DeepSeek

DeepSeek was founded in 2023 by Liang Wenfeng, who is also the founder of High-Flyer Capital Management, a hedge fund currently the only company funding DeepSeek. The company’s first product was DeepSeek Coder, an open-source model specialized in coding tasks, launched in November 2023.

Currently, the DeepSeek team is mainly composed of young individuals from top Chinese universities, focusing on technological innovation. Moreover, not having external investors due to its single funding model has allowed the company to evolve without pressure towards long-term AI projects.

However, DeepSeek is subject to evaluations by China’s Internet regulator, whose task is to ensure that this technology complies with “fundamental socialist values.” This implies that the AI does not respond to certain inquiries considered to be sensitive topics.

To show this, I asked it about two different topics: the Tiananmen Square protests of 1989 and also about the Spanish Civil War. And guess which of them was not answered. The AI assitant staed that it cannot respond to the first question, however, it does provide a summary about the second topic.

DeepSeek R1

On November 20, the Chinese startup presented a preview of its DeepSeek R1 AI, and on January 20, it released it as an open version. It is an open-source reasoning model whose performance, according to reports, exceeds that of o1 from OpenAI in certain parameters.

Specifically, the parameters where R1 excels are AIME (measures the efficiency of an AI model), MATH-500 (focused on mathematical problems), and SWE-bench Verified (programming tasks).

Being a reasoning model, R1 can verify itself and the responses it provides through deep thinking. This means that its response times are slightly slower than those of other generative AIs, but it guarantees more complete and accurate responses. DeepSeek R1 reasons through tasks, plans in advance, and executes actions that allow it to reach the correct result.

It has only been 3 days since Deepseek R1 was released and it is INSANE

SPOILER: ChatGPT is falling behind.

13 amazing examples so far (do not miss the fifth one) pic.twitter.com/QsHNWEmp7d

— Filipe | AI (@filicroval) January 24, 2025

This AI contains 671 billion parameters (internal variables used by AI models to make predictions or decisions), although the developer has released two reduced versions ranging from 1.5 billion to 70 billion parameters. This way, they can adapt to the system in which they will be executed, for example, the smallest version is suitable for a laptop, while the full version requires high-powered hardware.

Currently, the R1 model is available on Hugging Face under an MIT license and with no commercial restrictions.

DeepSeek V3

Just a few days after releasing R1 as an open version, DeepSeek presented its model V3. This AI is a MoE (Mixture-of-Experts) language model, which means it is based on an advanced neural architecture capable of dividing learning among various “experts” specializing in specific tasks or data subsets. This way, it combines the strengths of several specialized models to achieve superior overall performance.

DeepSeek V3 consists of 671 billion parameters, with 37 billion activated for each token. This model is capable of handling a variety of text-based tasks, such as coding jobs, translation, and writing, through descriptive prompts.

According to data published by the company itself, DeepSeek V3 surpasses models such as Llama 3.1 405B, Claude 3.5, or GPT-4 in various parameters.

Introducing DeepSeek-V3!

Biggest leap forward yet:
⚡ 60 tokens/second (3x faster than V2!)
Enhanced capabilities
API compatibility intact
Fully open-source models & papers

1/n pic.twitter.com/p1dV9gJ2Sd

— DeepSeek (@deepseek_ai) December 26, 2024

Likewise, this AI has garnered significant attention because its development required 2,788 million hours of training and cost 6.05 million dollars. While this last figure is overwhelming, when compared to the GPT-4 model from OpenAI, it appears “modest,” since this model required an investment of around 88 million dollars.

V3 is released under a license that allows developers to download and modify it for a wide range of applications, including commercial apps. You can access this model on the DeepSeek website, on GitHub, or in its app.

Photo: DeepSeek and Canva