In the world of AI, there was a general concept that developing cutting-edge, large-scale language models requires important technical and financial resources. This is one of the main reasons why the US government has pledged to support the $500 billion Stargate project announced by President Donald Trump.
Listen to this article. This audio was generated by AI.
However, Chinese AI developer Deepseek has confused the concept. On January 20, 2025, Deepseek released the R1 LLM for just a small percentage of the costs incurred by other vendors in their own developments. Deepseek offers the R1 model under an open source license and is free to use.
Within days of release, Deepseek AI Assistant (a mobile app that provides the chatbot interface for DeepSeek-R1) is at the top of Apple's App Store chart and outrank Openai's ChatGPT mobile app. Investors questioned the value of large US-based AI vendors, including NVIDIA, leading to an increase in Deepseek meteors in terms of use and popularity that led to the sale of the stock market on January 27, 2025. Microsoft, Meta Platforms, Oracle, Broadcom and other tech giants also saw a significant drop as investors reevaluated their AI ratings.
What is Deepseek?
Deepseek is an AI development company based in Hangzhou, China. The company was founded in May 2023 by Liang Wenfeng, a graduate of Zhijiang University. Wenfeng also co-founded High-Flyer, a China-based quantitative hedge fund that owns Deepseek. Currently, Deepseek operates as an independent AI research lab under a high-end umbrella. The full funds and Deepseek ratings have not been made public.
Deepseek focuses on developing open source LLM. The company's first model was released in November 2023. The company has iterated multiple times with Core LLM and built several different variations. However, it did not become world famous until January 2025, after the release of the R1 inference model.
The company offers multiple services to its model, including web interfaces, mobile applications and API access.
Openai vs. Deepseek
Deepseek represents the latest challenge for Openai, which established itself as an industry leader with its ChatGPT debut in 2022. Openai helped advance the generation AI industry with models from the GPT family and O1 class inference models, including O3 and O4 Mini.
Both companies develop generative AI LLM, but there are a variety of approaches.
| Openai | deepseek | |
| Year of establishment | 2015 | 2023 |
| Headquarters | San Francisco, California. | Hangzhou, China |
| Development Focus | Wide range of AI features | Efficient, Open Source Model |
| key Model | GPT-4O, O1, GPT-5 | DeepSeek-V3, DeepSeek-R1, DeepSeek-V3.1 |
| Specialized model | Dall-E (image generation), Whisper (voice recognition) |
Deepseek Coder (Coding), Janus Pro (Vision model) |
| API Pricing (per 1 million tokens) |
GPT-5: $1.25 (input), $10 (output) | deepseek-v3.1: $0.56 (input), $1.68 (output) |
| Open Source Policy | limited | Mainly open source |
| Training Approach | Fine adjustments based on supervised instruction | Reinforcement learning |
| Development Cost | O1 hundreds of millions of dollars (estimated) |
According to the company, the DeepSeek-R1 is under $6 million |
Training Deepseek's innovation
Deepseek trains the R1 model using a different approach than OpenAI uses. Training time is less, fewer AI accelerators are less, and development costs are less. The purpose of Deepseek is to achieve artificial general information, and the company's advances in reasoning capabilities represent a major advance in AI development.
The research paper outlines several innovations that Deepseek developed as part of the R1 model.
- Reinforcement learning. Deepseek used a large-scale reinforcement learning approach focusing on inference tasks.
- Reward engineering. Researchers have developed a rule-based reward system for models that outweigh the more commonly used neural reward models. Reward engineering is the process of designing an incentive system that guides learning of AI models during training.
- distillation. Using efficient knowledge transfer techniques, DeepSeek researchers compressed the functionality into a small model of 1.5 billion parameters.
- Emergency Action Network. Deepseek's emergency action innovation is the discovery that complex inference patterns can naturally develop through reinforcement learning without explicitly programming them.
- Hybrid architecture training. Deepseek v3.1 introduces an evolved training methodology. Here, a single model is trained to support both fast inference and deep inference modes. This hybrid approach required developing new chat templates and tokenization strategies, including the introduction of specific thinking tokens (
and ) controls the inference behavior of the model.
Deepseek major language model
Since its founding in 2023, Deepseek has released a series of generator AI models. With each new generation, the company has been working to advance both the functionality and performance of the model.
- Deepseek coder. Released in November 2023, this is the company's first open source model designed specifically for coding-related tasks.
- Deepseek LLM. Released in December 2023, this is the first version of the company's generic model.
- deepseek-v2. Released in May 2024, this is the second version of the company's LLM, focusing on powerful performance and reducing training costs.
- deepseek-coder-v2. Released in July 2024, it is a 236 billion parameter model that provides a 128,000 token context window designed for complex coding challenges.
- deepseek-v3. Released in December 2024, DeepSeek-V3 uses a mixed architecture that can handle a variety of tasks. This model has 671 billion parameters with a context length of 128,000.
- deepseek-r1. Released in January 2025, the model is based on the DeepSeek-V3 and focuses on advanced inference tasks that compete directly with OpenAI's O1 model, while maintaining a significantly lower cost structure. Like DeepSeek-V3, this model has 671 billion parameters with a context length of 128,000.
- DeepSeek-R1-0528. The R1-0528 model released in May 2025 is an updated version of the original R1 model. This model supports system prompts, JSON output, and function calls, making it more suitable for agent AI use cases. Deepseek also claims that hallucination rates are lower and more accurate compared to previous releases. R1-0528 also benefited from a large inference depth of 12,000 tokens on average in previous versions and an average of 23,000 tokens per question.
- deepseek-r1-0528-qwen3-8b. A smaller, distilled version based on Alibaba's QWEN3 model aimed at systems with limited computing resources. According to DeepSeek, this 8 billion parameter model is consistent with the performance of the larger QWEN3-235B model.
- Janus-Pro-7b. Released in January 2025, the Janus-Pro-7B is a vision model that allows you to understand and generate images.
- DeepSeek-V3.1. Released in August 2025 as a hybrid model with dual mode features, DeepSeek-V3.1 supports both thought modes and non-think modes within a single model. The model is built on an 840 billion parameter base and supports 128K context lengths. This model also supports enhanced tool invocation and agent functionality through post-training optimization.
Why are they raising alarms in the US?
There was a lot of hype around the release of DeepSeek-R1, but it caused alarms in the US, causing concern and stock market sales in high-tech stocks. On Monday, January 27th, 2025, NASDAQ Composite fell 3.4% when market opening, while Nvidia fell 17%, losing roughly $600 billion in market capital.
Deepseek has raised alarms in the US for several reasons, including:
- Cost confusion. Deepseek claims to have developed the R1 model for less than $6 million. Low-cost development threatens the business model of US high-tech companies that have invested billions in AI. Deepseek is cheaper for users than Openai.
- Technical achievements despite limitations. Exports of top-performing AI accelerators and GPU chips from the US are limited to China. But nonetheless, Deepseek has demonstrated that cutting-edge AI development is possible without access to the most advanced US technologies.
- Business mOdell's threat. In contrast to Openai, a proprietary technology, DeepSeek is open source and free, challenging the revenue model for US companies that charge monthly fees for AI services.
- Geopolitical concerns. China-based Deepseek is challenging US technological domination in AI. Tech investor Mark Andreessen called it AI's “Sputnik Moment” and compared it to a breakthrough in the Soviet Union's alien race in the 1950s.
Deepseek Bans
Countries and organizations around the world have already banned ethics, privacy and security issues within businesses. As all user data is stored in China, the biggest concern is the possibility of data leaks to the Chinese government. LLM was also trained in the Chinese worldview. This is a potential issue by the country's authoritarian government.
Where Deepseek is prohibited includes:
- Australian government agency.
- India's central government.
- Italy.
- NASA.
- Ministry of Industry in Korea.
- Taiwan government agency.
- Texas government.
- US Congress.
- US Navy.
- US Pentagon.
There is also the possibility of a widespread ban in the European Union. Several countries within the EU are taking action, including Italy and Germany. Italy banned Deepseek in January 2025 on data privacy concerns. Germany followed in June 2025 when the country's data protection commissioner Meike Kamp officially asked Apple and Google to consider blocking the Deepseek app from the German app store. The concern is that DeepSeek is transferring user data to China without proper protection under EU data protection regulations. Deepseek potentially violates several EU regulations, including general data protection regulations and EU AI laws.
In the case of DeepSeek, the EU's regulatory approach poses challenges regarding data sovereignty, transparency requirements, and the need to demonstrate appropriate safeguards of European user data.
DeepSeek CyberTack
Deepseek's popularity has not been noticed by cyberattackers.
On January 27, 2025, Deepseek reported a massive malicious attack on the service, forcing the company to temporarily restrict new user registrations. The timing of the attack coincided with Deepseek's AI assistant app, which overtook ChatGpt as the top downloaded app on the Apple App Store.
Despite the attack, DeepSeek maintained its service for existing users. The issue was extended until January 28th, reporting that the company has identified the issue and deployed fixes.
Deepseek does not specify the exact nature of the attack, but widespread speculation from public reports suggests that it is some form of a DDOS attack targeting APIs and web chat platforms.
Deepseek data has been released
Teams within Wiz Research-Cloud Security Vendor Wiz Inc. – The survey results were published on January 29th, 2025. The information includes DeepSeek chat history, backend data, log streams, API keys, and operational details. DeepSeek took the database offline immediately after being notified. It is unknown how long the database has been published.
Sean Michael Kerner is an IT consultant, technology enthusiast and Tinkerer. He is known to draw a token ring, configure netware, and compile his own Linux kernel. He consults with industry and media organizations about technology issues.
Read more DeepSeek's report:
Deepseek under the microscope: sanity check
deepseek-r1 easily generates more dangerous content than other major language models
AI jailbreak techniques have proven to be extremely effective against DeepSeek
Wiz reveals the API keys and chat history published by DeepSeek database
New artifacts extend observability to deepseek
Microsoft, AWS, and Cerebras launch the DeepSeek-R1 model
