DeepSeek: Everything You Need to Know

DeepSeek, a little-known Chinese AI startup, became the headline of all major media platforms. It has the power to compete with ChatGPT at a much lower computing cost for training. Its popularity caused a sudden drop in US stock pricing, showing its powerful impact among other major AI competitors. DeepSeek-R1’s performance surprised many people outside the country. China’s researchers claim that DeepSeek’s success makes the country a global leader in artificial intelligence (AI).

DeepSeek works similar to ChatGPT. It used OpenAI’s model to train its competitor using a distillation process. It helps businesses and individuals automate tasks, search data, and improve decision-making. It is a large language model that responds with highly accurate text-based answers. Its mobile app has 10M+ downloads to date. It can produce results in English and Chinese languages with its bilingual capabilities.

Let's explore what is DeepSeek, its statistics, how it works, its unique features, and its pros and cons.

What is DeepSeek?

DeepSeek was first introduced by Liang Wenfeng in May 2023. He is a prominent personality in both the AI world and the hedge. DeepSeek got operational funds from High-Flyer, a hedge fund launched by Wenfeng. It uses AI to analyze financial data to make investment decisions. This funding allowed long-term research and development of advanced models. Liang mentioned in his speech that:

If the US can develop its quantitative trading sector, why not China?

When asked about DeepSeek's success, he said:

Their surprise stems from seeing a Chinese company join their game as an innovator, not just a follower - which is what most Chinese firms are accustomed to.

Here is a step-by-step development and advancement of DeepSeek model:

• DeepSeek entered the AI market in November 2023 with DeepSeek Coder. It was an open-source AI model for coding.
• Then, DeepSeek LLM, comprising 67B parameters, was introduced in December 2023. It could compete with other LLMs.
• In May 2024, the company launched DeepSeek-V2 at a much lower cost and great performance. Its affordable rates forced other tech giants to lower their rates. This approach help them to stay competitive.
• DeepSeek-Coder-V2 was then released in July 2024 with 236 billion parameters. It can do complex coding tasks and write lengthy content up to 128k tokens.

Finally, the company launched its latest models, DeepSeel-V3 and DeepSeek-R1. These have much higher accuracy and authenticity.

• DeepSeek-V3 comprised of 671 billion parameters was launched in December 2024. It required the least resources compared to others.
• DeepSeek-R1 released in January 2025 is excellent at performing reasoning tasks. R1 model is a “chain of thoughts” as it logically builds responses and explains them step-by-step for every query. It works similarly to the human mind in reasoning and sorting answers.
• Janus-Pro-7B is a vision model. Itwas released in January 2025 to understand and generate photos.

The company also released several distilled models based on synthetic data generated by R1 like Llama and Qwen. These DeepSeek-R1-Distill differ in their efficiency for varying hardware configurations and computational demands.

Workforce at DeepSeek Company

The founder of DeepSeek company was involved in the research process. He preferred local talent over workers who studied overseas for the company’s campus in Hangzhou. The exact number of the company’s workforce is unknown. Liang Wenfeng told that he recruited talented gaduates and doctoral level students. Some of them are under 35 and well-known as tech superpowers in China. He complimented them as:

They are deeply motivated by a drive for self-reliance in innovation.

The company pays a handsome amount to the recruited top talent. They attract developers from companies to join their workforce with higher pays than usual. The company offers jobs at the Chinese recruitment site, Zhipin up to 60,000 yuan monthly. It is comparatively double what is offered by other tech giants. For instance, the average salary at Tencent is about 35,000 yuan monthly.

Key Statistics of DeepSeek

A quick overview of DeepSeek statistics here:

• DeepSeek spent only $5.6 million to train the model for extraordinary outperformance that’s
  10% of Meta’s Llama.
• The website has 33.7 million monthly active users globally as of January 2025.
• DeepSeek got 22.15 million daily active users globally as of January 2025.
• The company receives 51.24%of its traffic from China, India, and Indonesia.

How Does DeepSeek Work?

DeepSeek used a different way to train its R1 models, unlike OpenAI. The company’s focus was on using fewer AI accelerators at a lesser cost and in the least time. DeepSeek developed several innovations as a part of its R1 model. These are:

• Reinforced Learning: Large-scale reinforcement learning approach for reasoning tasks.
• Reward Engineering: Rule-based Reward System to outperform neural reward models.
• Distillation: Compression of capabilities as small as 1.5b parameters into models.
  Emergent Behavior Network: The complex reasoning patterns can develop this behavior without programming.

DeepSeek uses a Mixture-of-Experts (MoE) method to make its AI model efficient and powerful. MoE further divides the tasks among smaller models known as “experts.” Each expert is specialized to perform a special task. This approach reduces computing power requirements for more efficient AI development.

Key Features: What Makes DeepSeek Unique?

DeepSeek comprises several advanced options that make it unique among its top rivals. Some of the most prominent features include:

1. Extraordinary Models
DeepSeek-V2 and DeepSeek-V3 are valuable for several applications. They share highly accurate results at an exceptional rate.

2. Cost-Effective
DeepSeek provides advanced AI services at affordable rates. It makes it highly economical among its competitors. The tool costs only from $0.27 to $1.1 per million tokens. Being less energy-intensive also makes it environment-friendly for emissions. It's also a step forward towards sustainable AI scaling.

3. Affordable API Pricing
DeepSeek API pricing is also lower than its competitors. Its R1 API costs $0.55/million input and $2.19/million output tokens. These figures are lower than OpenAI’s API costing $15/million for input and $60/million for output tokens.

4. Bilingual
Its software is trained in two languages including English and Chinese to enhance its user base.

5. Open-Source Code
DeepSeek’s open-source accessibility makes it customizable for users. They can change the code and integrate it into their businesses.

6. Lengthy Content
DeepSeek supports content lengths up to 128,000 tokens. It can do complex analyses and provide extensive responses with this power.

China’s Reaction to DeepSeek’s Impact

Silicon Valley and Wall Street giants are shaken up with the abrupt popularity of DeepSeek-R1. Marina Zhang, an associate professor at the University of Technology Sydney says:

In China, DeepSeek's advances are being celebrated as a testament to the country's growing technological prowess and self-reliance. The company's success is seen as a validation of China's Innovation 2.0, a new era of homegrown technological leadership driven by a younger generation of entrepreneurs.

Wei Sun, principal AI analyst at Counterpoint Research says.

DeepSeek has proven that cutting-edge AI models can be developed with limited compute resources. In contrast, OpenAI, valued at $157 billion, faces scrutiny over its ability to maintain a dominant edge in innovation or justify its massive valuation and expenditures without delivering significant returns.

Pros and Cons of DeepSeek

Pros

• Open source
• Time saver with its automation power.
• Easy to explore.
• Free to use for general users and even developers.
• Provides comprehensive reports of PDFs.
• Helpful for better teamwork.

Cons

• Might be pricy for startups or small teams.
• Take time to learn.
• Privacy issue as the app is linked with a Chinese government mobile firm and shares data with them.

Why Countries are Banning DeepSeek?

The software stores information which raises questions about how it handles personal data. DeepSeek has a hidden code in its programming that sends users data to the Chinese government. The application sends information to a telecommunication company, CMPassport.com, an onlineregistry for China Mobile. OpenAI also collects personal information according to ChatGPT’s privacy policy. Eddy Borges-Rey, associate professor in residence at Northwestern University in Qatar said to Al Jazeera:

Virtually all major tech companies – from Meta to Google to OpenAI – exploit user data to some extent. They use data for targeted advertising, algorithmic refinement and AI training. Many have been fined or investigated for privacy breaches, but they continue operating because their activities are somewhat regulated within jurisdictions like the EU and the US.

He added:

Chinese platforms such as DeepSeek are treated differently by the West because they are seen as operating under the jurisdiction of the Chinese government, which has laws [such as the National Intelligence Law] that theoretically allow state access to corporate data.

Rep. Josh Gottheimer (D-NJ), who serves on the House Intelligence Committee, told ABC News.

I think we should ban DeepSeek from all government devices immediately. No one should be allowed to download it onto their device. And I think we have to inform the public. DeepSeek's terms of service specify that they "shall be governed by the laws of the People's Republic of China.

Which Countries Banned DeepSeek Application?

Australia
Australia has banned DeepSeek on government devices for security risks at national level. Australian government says DeepSeek is “absolutely not a symbolic move” and banned it. DeepSek products are of high risk for the country according to Australian government. Hence, they prohibited the use of DeepSeek among their systems and smart devices.

Canada
The Canadian government showed “serious privacy concerns” and restricted its use. It retains sensitive personal data that makes it inappropriate to install.

Italy
Italy blocked DeepSeek app and ordered the company to not process their citizen’s data. The country’s data protection authority restricted use of DeepSeek for storing details.

Netherlands
Netherlands banned civil servants from using DeepSeek at workstation. The country doesn’t want to share sensitive information with Chinese servers. The Dutch Data Protection Authority also warned residents to use app with caution.

South Korea
The government asks DeepSeek about how it handles user data but the company fails to respond. Thus, South Korean authorities blocked DeepSeek’s access on computers. According to South Korea’s company, DeepSeek trained itself on collecting personal data.

Taiwan
DeepSeek imposes serious security concerns as China collects personal dats at their end. The country’s digital ministry recommends government departments to not use the application.

United States
Washington lawmakers passed a bill to ban DeepSeek on government devices. They mentioned it as “alarming threat to US national security.” According to CNBC report, US NASA also blocked platform access from its system. Moreover, US Navy restricted DeepSeek use due to “potential ethical and security concerns.”

Conclusion

AI market is volatile which makes it challenging to predict its future. DeepSeek works on AGI principle that can think and respond to queries like humans. There are several risks associated with AGI implementation. One of the major concerns is the loss of control over systems humans create. What advancement will DeepSeek’s R1 bring to AI sector is a time-dependent factor. To stay competitive, DeepSeek must improve itself with exceptional advancements in the system. It requires consistent improvement and introduction of unique capabilities. DeepSeek has the potential to reshape future with its commitment to advance AI.

Do you know any startup that can compete DeepSeek now? Mention it in the comment section. So, we can find where DeepSeek lacks and explore other AI chatbot options too.