What Is Deepseek?

The United States has long been seen as the dominant player in the global race to control artificial intelligence — that is, until the emergence of DeepSeek. The Chinese AI company first turned heads with the release of its DeepSeek-R1 model in January 2025, which rivaled the performance of larger American models at a fraction of the cost. In response, tech stocks plummeted as U.S. investors began to question whether the billions of dollars being poured into AI infrastructure and related initiatives were really worth it.

American tech companies have refused to give up their advantage, though, ramping up their infrastructure investments to fend off DeepSeek and other Chinese challengers. With the stakes higher than ever, we’re taking a closer look at DeepSeek’s history, its current slate of models, how they work and the pros and cons they offer. 

More on the U.S.-China AI RaceA Guide to How America Controls Nvidia Chip Exports to China

What Is DeepSeek? 

DeepSeek is a Chinese AI startup that was founded in 2023 by Liang Wenfeng, a technology entrepreneur with a background in engineering who also founded the quantitative trading firm High-Flyer Capital Management. High-Flyer became DeepSeek’s main funding source, initially using AI models to guide investment decisions before establishing the AI research lab to achieve artificial general intelligence (AGI), or AI that can think and learn like humans. 

The lab is known for developing open-weight models of the same name. The term “open weights” refers to a model’s finalized parameters that are released without the original training data or code. This way, users cannot copy the model exactly or fully understand its training process, but they can still tailor a public version to their preferences and run it on their own machines.

DeepSeek is also the name of the company’s AI assistant, available on chat.deepseek.com and as a mobile or desktop application. DeepSeek’s latest models power this AI tool, giving consumers access to enhanced capabilities that can further personalize everyday conversations and interactions. 

DeepSeek Models 

DeepSeek has released a series of game-changing models over the years. Here are a few of the most impactful additions to its lineup. 

DeepSeek-V3

DeepSeek-V3 is a fully open-source model with 671 billion total parameters and can process and generate responses at 60 tokens per second, triple the speed of earlier DeepSeek models. It demonstrates strong language understanding in English and Chinese and scores highly in coding, keeping pace with other open- and closed-source models despite needing less training time. That said, V3 was trained on 14.8 trillion high-quality tokens. 

DeepSeek-R1

The latest version of DeepSeek-R1 has advanced reasoning abilities, using 11,000 more tokens per query and increasing its accuracy by 17 percent compared to the previous version. A major driver of R1’s reasoning is its “thinking mode” that lets it review its initial response and refine its accuracy. By using post-training techniques to get the most out of its algorithms, R1 performs at a high level while limiting hallucinations.  

DeepSeek-V3.1

DeepSeek-V3.1 represents DeepSeek’s first meaningful step toward agentic AI, or AI that can solve complex, multi-step tasks without constant human guidance. It’s trained on 840 billion tokens, allowing it to retain an extensive context of up to 128,000 tokens in both thinking and non-thinking mode. As a result, V3.1 greatly improves upon R1’s ability to answer short queries and browse the internet for information. 

DeepSeek-V3.2

DeepSeek-V3.2 marked another reasoning milestone for DeepSeek, rivaling OpenAI’s GPT-5 model at the time. It was trained on data from more than 1,800 environments and more than 85,000 instructions, so it can direct AI agents to use different tools. The company also introduced a new mechanism called DeepSeek Sparse Attention, which simplified V3.2’s computational process while still enabling quality outputs in response to longer queries.   

DeepSeek-V4

The initial preview version of DeepSeek-V4 teased a context length of 1 million tokens and a slate of V4 models. For instance, V4-Pro has 1.6 trillion parameters, giving it the compute to become a top model at solving math and coding problems. It performs particularly well on agentic coding benchmarks and possesses extensive real-world knowledge, falling short only of Google’s Gemini 3.    

DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash is the smallest model in the V4 family, activating as few as 8 billion of its 552 billion parameters to process inputs. The model cuts down on storage to reduce usage costs, yet it holds its own on programming, cybersecurity and automation benchmarks. Perhaps the secret behind its efficient performance is its emphasis on post-training reinforcement learning (RL), or training a model to finish a task through trial and error.

Another Chinese AI Company to WatchChina’s New Kimi K3 Model Has the American AI Industry on High Alert

How Does DeepSeek Work? 

DeepSeek uses large language models, systems trained on massive amounts of data to process and produce human language. Below are the typical training stages for LLMs that DeepSeek’s models follow:   

• Pre-Training: To learn human language, a model labels its own training data gathered from public and licensed sources. This process is called self-supervised learning, which combines elements of supervised learning — training a model on labeled data — and unsupervised learning — training a model on raw, unlabeled data. 
• Supervised Fine-Tuning: The pre-trained model is then trained on smaller data sets with labeled examples. This approach tailors the model to specific use cases and allows it to apply its linguistic knowledge from pre-training. 
• Reinforcement Learning: By learning through trial and error, DeepSeek’s models can remember past mistakes, adjust their responses and continually improve how they handle different situations. Researchers may intervene during this stage to ensure outputs are accurate and appropriate.  

Another key technique DeepSeek uses is model distillation, or training a smaller model on a more complex one to transfer the large model’s capabilities to the smaller one. For example, DeepSeek-R1 was trained on open-weight models from the Llama and Qwen families, but it mainly distilled the larger DeepSeek-V3, which uses a mixture-of-experts architecture that activates only a small set of parameters for the model to operate at full capacity when answering a query. That’s the secret behind DeepSeek’s efficiency. 

So, a DeepSeek model processes a natural-language request by breaking the text down into tokens, sending the tokens to the relevant “experts” in its neural network, compiling data from outside sources as part of retrieval-augmented generation and using AI inference to draw insights from this unfamiliar data and inform its responses. For added transparency, users can ask DeepSeek to activate chain-of-thought reasoning, where the model walks a user step-by-step through its thinking.  

How Is DeepSeek Used? 

Given their multimodal nature, DeepSeek’s models can complete a range of tasks across various industries, including the following: 

• Customer Support: DeepSeek can hold basic conversations with customers 24/7 and answer common questions in more than 100 languages. 
• Software Development: Software developers can use DeepSeek to generate and review code and quickly patch bugs. 
• Data Science: DeepSeek can provide business intelligence insights by conducting market research, generating data visualizations and explaining dashboard findings. 
• Finance: DeepSeek can create financial reports and long-term forecasts, while detecting anomalies that may signal fraud. 
• Marketing: Marketers can bounce ideas off DeepSeek during brainstorming sessions and have it produce social media posts, web content and other marketing materials.  
• Education: Teachers can lean on DeepSeek to help design personalized curricula, and academics can use it to conduct in-depth research. 

Benefits of DeepSeek 

DeepSeek’s models are known for delivering high-quality outputs that require minimal compute, yielding several advantages: 

• Increased Efficiency: Because they need only a few parameters to address each query, DeepSeek models require far less compute during training and deployment.   
• Lower Costs: It costs much less to train DeepSeek compared to other models, resulting in much lower costs for users.  
• Improved Accuracy: DeepSeek breaks down queries into steps and reasons through complex questions to provide more accurate answers.   
• Coding Performance: DeepSeek’s advanced reasoning makes it ideal for coding scenarios, especially since it can explain its thinking to developers. 

DeepSeek has arguably changed the wider AI industry for the better as well by creating open-weight models anyone can use and customize for free. This decision has pressured other companies like OpenAI to release their own free open-weight models, making their technology more accessible for innovative research and everyday use.

The Next Phase in the Global AI RaceEurope Pushes Back on Trump’s AI Plan

Limitations of DeepSeek

Although DeepSeek’s models are cost-efficient, they come with some inherent flaws that users need to consider: 

• Algorithmic Bias: Like any AI tool, DeepSeek remains vulnerable to biases baked into its training data, sometimes leading to hallucinations.   
• Security Issues: Companies must establish their own security processes when using DeepSeek’s open-weight models, or risk their misuse.   
• Limited Availability: DeepSeek models aren’t integrated with common American tools like OpenAI’s ChatGPT, Anthropic’s Claude and Google’s Gemini.    
• Technical Expertise: Running DeepSeek locally requires teams to set up their own infrastructure, especially for larger models that demand more resources.

The AI race between the United States and China has also impacted access to DeepSeek models. American tech companies have accused DeepSeek of illegally distilling their models, despite the company’s explanation that the number of AI-generated outputs on the web likely led to unintentional distillation. Meanwhile, U.S. politicians have pushed to ban DeepSeek models on government devices. With AI becoming a matter of national security, DeepSeek could face fierce backlash in the U.S. and other countries for the foreseeable future.

Frequently Asked Questions