DeepSeek, a new AI powered by a Chinese company, has stirred heated discussions globally. While some people treat it as a healthy competition, others are concerned about problems with data privacy that appeared shortly after its release. These mixed reactions reflect broader questions about the future of international AI development and regulation.
Will DeepSeek be a serious alternative to ChatGPT and Gemini, or will it be buried by European data protection laws and US restrictions? In this article, we’ll go through the background of DeepSeek, its capabilities compared to other AI tools, user reception and controversies, and its business potential.
DeepSeek AI history
High-Flyer, the Chinese company behind DeepSeek was founded in December 2023 by Liang Wefeng, who has a background in implementing artificial intelligence for financial data analysis and investment decisions. In April 2023, the company created a lab dedicated to artificial intelligence and AI tools research. This has led to the development of large language models, which eventually helped to build the first chatbot app powered by DeepSeek-R1 technology.
In 2024, High-Flyer released its first assistant. The chatbot entered the global market on January 20, 2025, and quickly gained traction, provoking an international discussion about the technology.
DeepSeek capabilities
DeepSeek mobile app was released simultaneously for iOS and Android and rose in popularity, becoming the most downloaded application on the Apple App Store in the US at that moment. The tool, which is also available through web browsers, offers a familiar interface similar to other AI assistants.
What can DeepSeek be used for?
- Research and information gathering. The chatbot can collect data about a specific industry, summarize it, find relevant articles, and verify facts or statistics.
- Writing and editing. DeepSeek can help write content, edit and improve grammar or style, and generate ideas for social media posts and marketing campaigns.
- Data analysis and insights. The model can analyze data, identify trends, and generate relevant insights; it can propose how to present data and automate repetitive tasks.
- Project management. It can create to-do lists, set priorities, and track progress; it can also summarize meetings, take notes, and draft communication for teams.
- Learning assistance. DeepSeek can help create guides, tutorials, and explanations on how to use different tools, as well as assist in translating or learning languages.
- Coding help. The model can help debug, write code snippers, and explain programming concepts.
Looking at the scope of these capabilities, it’s clear that DeepSeek can compete with other chatbots with similar functionality. This is not the full list, as it can help also in many other areas, such as troubleshooting and strategic planning.
DeepSeek-V3 AI large language model architecture
DeepSeek-V3, an improved open-source Mixture-of-Experts (MoE) language model, has been developed for cost-effective training and efficient inference. The V3 model has 671 billion parameters, with 37 billion being activated per token, resulting in tremendous efficiency while maintaining remarkable performance across many benchmarks.
The model is trained using the HAI-LLM framework, which was developed internally. It uses optimizations like zero-bubble pipeline parallelism, expert parallelism, mixed-precision training with FP8, and multi-token prediction training. These aim to reduce communication overhead and improve computational efficiency.
This LLM enables extended context lengths of up to 128K tokens, addressing the issue of long-context processing in language models. Its key architectural features are:
- Multi-Head Latent Attention (MLA): MLA improves the classic Multi-Head Attention (MHA) algorithm to overcome inefficiencies resulting from the Key-Value (KV) cache during inference.
- Native Sparse Attention (NSA): Advanced attention mechanism enhancing the efficiency of long-context sequences processing in language models by employing a dynamic hierarchical sparse strategy, which combines coarse-grained token compression with fine-grained token selection.
- DeepSeekMoE Architecture: The architecture is specialized in sparse computation by dividing experts into smaller groups and isolating shared experts. It allows for more efficient training and parameter use than typical Mixture-of-Experts systems.
- Training Efficiency: DeepSeek-V3 has a 3x increase in token generation speed, reaching 60 tokens per second, an optimized inference throughput, and lower training cost, which makes it highly scalable for deployment.
All details of the models are available on DeepSeek’s Git Hub.
Performance evaluation and training infrastructure
DeepSeek-V3 excels across a variety of benchmarks. In English, it performs similarly to other open-source models, such as LLaMA, in reading comprehension, reasoning, and commonsense knowledge. In terms of code and math tests, it shows better results.
While these tests are conducted to verify the model’s capabilities on the same level as the others, they don’t translate directly to actual performance when implemented for specific tasks.
Still, DeepSeek-V3 is an innovative LLM that combines multiple benefits in a single package. It uses GPU resources with exceptional efficiency during the training and inference phases, making it a cost-effective deployment alternative. Moreover, its open-source nature assures full openness of both the model and its methodology, encouraging cooperation and innovation within the research and development community.
DeepSeek compared to other models
Today, users can choose from various generative AI chatbots tailored for different tasks, such as content creation, coding, or data analysis. Each tool has its strengths and limitations that should be considered when used for specific purposes.
| Area | DeepSeek | ChatGPT | Claude | Gemini |
| Training method | Task-specific processing | General-purpose training | Safety-focused AI training | Multimodal AI (text, image, video) |
| Response speed | Faster in niche tasks | Consistent across tasks | Slower but detailed responses | Fast and optimized for search |
| Response accuracy | Strong in technical fields | Better at complex queries | Strong in legal and policy topics | Efficiency and speed of responses |
| Best used for | Coding, technical, research | Creative writing, general research | Legal, ethical, structured tasks | Search-related tasks, creative work |
| Areas of underperformance | Struggles with creativity and general chat | Less specialized in niche technical fields | Lacks real-time internet updates | Long-context processing accuracy |
| Optimization options | Highly customizable | Limited customization | Strong emphasis on alignment | Adaptable, integrates with Google |
| Cost-structure | Free but data-sensitive | Free and subscription options | Free and enterprise plans | Free with advanced features in paid plans |
| Strengths | Efficient, open-source | Versatile, widely used | Safe, detailed reasoning | Multimodal capabilities, real-time search |
| Weaknesses | Biased, ethical concerns, data collection concerns | Outdated, biased, data concerns | Conservative, avoids risks due to AI safety-first approach | Can hallucinate, over-reliant on Google |
| Web access | Limited or no direct web access, internet search functionality not always operational | Integrated web browsing capabilities with access to real-time information | No direct web access, responses are based on its training data | Limited real-time web search, browsing capabilities not extensively documented |
Considering the privacy issues, Claude is an excellent choice regarding safety. It performs well in content-related tasks but is slower than other tools and doesn’t have web access. ChatGPT is the most versatile option with possible internet access, ensuring GDPR and CCPA compliance. Gemini offers multimodal capabilities but may be overreliant on external data sources. DeepSeek performs great in maths and reasoning tasks but lacks proper data security.
Business perspective
As already mentioned, the public noticed the open-source nature of DeepSeek’s model, which offers new possibilities for businesses that want to implement AI for specific purposes. The technology has become a significant competitor for Meta’s Llama 3 and OpenAI’s API. Moreover, the considerably lower price is something the other tech giants may have to worry about.
Yet, every model has different customization capabilities, performance, infrastructure, and costs. That’s why it’s necessary to analyze them all to find the right fit.
| Area | Meta’s Llama 3 | OpenAI’s GPT-4o | DeepSeek’s R1 |
| Customization | Fully open-source and allows for extensive customization. | A closed-source model with limited customization options. | Fully open-source and allows for extensive customization. |
| Fine-tuning | Offers fine-tuning the model for specific applications. | Offers fine-tuning through API with certain constraints. | Supports fine-tuning for specialized tasks. |
| Performance | Supports a broad range of applications. Available in 8B and 70B parameter versions. | Accepts both text and image inputs, producing text outputs. Capable of complex reasoning across multiple modalities. | Excels in reasoning, mathematics, and coding tasks. Comprises 671B parameters with 37B activated per query. |
| Infrastructure needs | Requires substantial computational resources for training and deployment, as well as significant technical expertise for effective use. | Accessible via OpenAI’s API, reducing the need for extensive local infrastructure; designed for straightforward integration into apps. | Developed with a focus on cost-effective training, utilizing fewer resources and available for local deployment. |
| Costs | Varies depending on the model, data used, and region. For the Llama 3.2 3B model, inference costs are $0.06 per 1 million input and output tokens. Training costs depend on the service provider. | $25 per million training tokens. Inference costs are $3.75 per million input tokens and $15 per million output tokens for the GPT-4o model. | Input at $0.14 per million tokens and output at $2.19 per million tokens for deepseek-reasoner. |
| Licensing | Commercial use above a certain threshold requires an additional license from Meta. | Fine-tuning is permitted under specific conditions. | Redistribution must include original copyright and permission notices. |
Llama 3 is a great choice for enterprises that require deep customization options. OpenAI’s GPT-4o offers ease of integration and great performance, while DeepSeek’s R1 is cost-effective and excels at complex reasoning and mathematics.
For those who would like to test DeepSeek, the best recommendation is not to train the model using real data or use only safe, neutral information. To do so, according to Klarna’s analytics manager, businesses should avoid sensitive inputs and personal tokens, monitor account activity, and use VPN/encryption. Contact us to get a safe plan for introducing GenAI into your business.
It’s also possible to test the DeepSeek-R1 model safely using platforms like Hugging Face or Perplexity AI, which ensure the data used is not shared with external entities. The services can host the model without Chinese government access restrictions or the need for VPNs.
If you’re curious about a specific case of fine-tuning DeepSeek’s R1 model, here’s an example of how it was done for Q&A tasks.
See examples of how GenAI can be safely implemented into banking and business operations.
The reception of the GenAI chat
Within just a few days of DeepSeek’s launch, users who tested it pointed out several visible benefits:
- People noticed lower costs of model fine-tuning compared to GPT. The API pricing for a million input tokens is $0.14 for DeepSeek and $2.50 for ChatGPT-4o.
- Advanced capabilities (Multi-Head Latent Attention) for generating content, coding, and solving problems.
- Open-source nature allows companies to fine-tune models, which is a great asset for many industries that need the processing of complex technical tasks.
Users also noticed some problems with privacy and censorship that were widely commented on across the web.
However, DeepSeek has done something many expected from American tech giants—made their AI model freely available for worldwide use, especially taking into consideration the announcement of publishing 5 code repositories for full transparency. While OpenAI allows only some of the models to be fine-tuned through their API, DeepSeek-R1 is completely open-source, giving users full access to adapt it as needed.
Privacy concerns and security risks
One of the key issues with the chatbot is DeepSeek’s data collection practices. Users have raised alarms, particularly concerning its policy of storing user data on servers located in China. As a result, due to local regulations, there is potential for state access to data stored in the Chinese resources as local companies must cooperate with state intelligence. Since the data is not protected by GDPR or other regulations, users can’t demand to see what information is stored about them and request it to be deleted.
NowSecure’s assessment of security and privacy detected many issues in the DeepSeek iOS app that had already been downloaded by over 2.5 million users. The company enumerated multiple serious privacy concerns, such as unencrypted data transmission, insecure data storage, extensive data collection, and more.
Theori Security Assessment published a similar summary of DeepSeek’s hidden risks. The experts discovered vulnerability to jailbreaks, meaning it’s easy to bypass safety filters and gain instructions for illicit activities, like malware creation. Theori experts also described security lapses, such as an exposed database that was accessible without authentication. While DeepSeek quickly responded to these incidents, new security and privacy concerns might still arise in the future. This is why many governments have taken action to prevent access to the tool.
The South Korean National Intelligence Service also commented on these issues, claiming that the app excessively collected personal data. After this discovery, the Ministry of Industry temporarily stopped government employees from using it. Soon after, the Personal Information Protection Commission blocked new downloads of DeepSeek on the App Store and Google Play in the country.
Shortly after the release of the AI powered by the Chinese organization, on 30 January 2025, Italy’s Data Protection Authority blocked DeepSeek. The reason? When a company offers services to EU citizens, it must comply with GDPR rules. DeepSeek violated these regulations by transferring user data to China without proper safeguards and having an unclear privacy policy that lacks information about data retention and user rights.
Italy was not the only country that has taken steps to limit access to DeepSeek. Australia has almost blocked the tool but, in the end, warned critical infrastructure operators not to use it for security reasons. Taiwan and India advised public sector departments against using this technology.
In the USA, lawmakers have proposed a bill to prohibit DeepSeek’s use on government devices. But the chatbot was not the only tool subject to US restrictions. For example, TikTok, another Chinese app, faced bans due to national security and data privacy concerns.
These actions were taken to ensure the safety of data and information of EU citizens and governmental bodies in other countries. However, DeepSeek responds to reports of risks, like the accidental exposure of the database detected by Wiz Research, by introducing proper security measures.
Still, users need to be cautious about using any AI tools. It’s better to avoid uploading any sensitive data or information into the chats and have limited trust in the privacy practices of all AI solution providers.
Censorship controversies
One of the first allegations against DeepSeek was the censorship of its own replies live when inquired about the Chinese government, such as the Tiananmen Square events and the status of Taiwan. The internet was full of examples of how the model refused to discuss these subjects.
Here is a video of DeepSeek censoring itself live when asked about the freedom of speech in China. It shows the process of the model generating the response and then changing it to “Sorry, that’s beyond my current scope. Let’s talk about something else”.
People came up with many ways to go around censorship:
There are other ways to bypass the AI’s defenses, including asking forbidden questions and requesting an answer with letters in words separated by a dot.
When ChatGPT was released, people also tried to go over safeguards to get information about how to build a bomb or other ways to harm people that go against safety policy. While early attempts to make ChatGPT roleplay these answers often succeeded, OpenAI has since strengthened its defenses. Given DeepSeek’s prompt reaction to the reported data exposure, it seems they will soon safeguard such censorship bypasses, too.
Intellectual property dispute
DeepSeek is not the only model that has stirred discussion about intellectual property. Gemini has also faced legal problems because of IP rules. As it turned out, Google trained its model on content from publishers and news agencies without notifying them. As a result, the company was fined $272 million for the breaches. A similar lawsuit was filed by the New York Times against ChatGPT and other chatbots.
DeepSeek faces an intellectual property theft allegation from OpenAI. The US tech giant accused the Chinese company of using their models to create a competing product. OpenAI claims High Flyer used the “distillation” technique that allows small AI models to be trained to replicate the behavior of a larger model by learning from its outputs without authorization.
The distillation technique is a legal practice, but according to OpenAI’s terms of service, it cannot be used to create a competitive AI system. What can it be used for? Distillation is adopted to fine-tune open-source models for specific tasks. It was implemented by researchers from Stanford University and the University of Washington to train the s1 model using OpenAI’s technology that can double-check its answers.
Final thoughts
DeepSeek has added new choices to the AI world. It opened up new opportunities for businesses in terms of model training for specific purposes. DeepSeek’s technology has made it available to more individuals thanks to its lower cost and ease of access, democratizing the AI market. The release of these open-weight AI models may also prompt the US tech giants to adopt more open-source practices and lower their prices to compete with DeepSeek.
Yet, with all the issues concerning data privacy and storage, feeding it with company data should be well thought-through. At this moment, the tool is too young as a technology to be safely used. Looking at the independent tests, DeepSeek still has a lot to do to cover all security problems.
FAQ
What’s the difference between Perplexity and DeepSeek?
These two tools are both AI-powered but have different purposes. Perplexity AI is a conversational search engine with access to web data that can look for fact-based answers. DeepSeek is a chatbot designed to handle technical tasks such as code generation.
Sources
https://www.nextbigfuture.com/2025/01/deep-deepseek-history-and-impact-on-the-future-of-ai.html
https://www.euroconsumers.org/the-full-story-of-deepseek-how-euroconsumers-is-driving-action-for-consumers/
https://en.wikipedia.org/wiki/DeepSeek
https://www.siliconrepublic.com/business/google-france-fine-alphabet-ai-bard-gemini-news-content
https://www.nowsecure.com/blog/2025/02/06/nowsecure-uncovers-multiple-security-and-privacy-flaws-in-deepseek-ios-mobile-app/
https://www.linkedin.com/pulse/deepseek-privacy-analysis-ignacio-masdeu-nms6f/
https://www.reddit.com/r/singularity/comments/1id60qi/big_misconceptions_of_training_costs_for_deepseek/?rdt=45300
https://zapier.com/blog/deepseek-vs-chatgpt/
https://www.datacamp.com/blog/deepseek-vs-chatgpt
https://aibusiness.com/nlp/france-fines-google-272m-for-training-ai-on-news-articles
https://www.reuters.com/technology/artificial-intelligence/south-koreas-industry-ministry-temporarily-bans-access-deepseek-security-2025-02-05/
https://time.com/7210875/deepseek-national-security-threat-tiktok/
https://nypost.com/2025/01/29/business/openai-says-it-has-proof-deepseek-used-its-technology-to-develop-ai-model/
https://www.bbc.com/news/articles/c5yv5976z9po
https://www.theguardian.com/technology/2025/jan/28/we-tried-out-deepseek-it-works-well-until-we-asked-it-about-tiananmen-square-and-taiwan
https://www.reuters.com/technology/amid-deepseek-frenzy-chinese-companies-detail-use-ai-2025-02-09/
https://www.theguardian.com/technology/2025/jan/28/chinese-ai-chatbot-deepseek-censors-itself-in-realtime-users-report
https://nypost.com/2024/01/15/business/openai-in-tough-spot-with-news-groups-as-it-faces-copyright-lawsuits-experts/
https://www.reuters.com/technology/artificial-intelligence/south-korea-spy-agency-says-deepseek-excessively-collects-personal-data-2025-02-10/
