DeepSeek is the talk of the world right now.
The AI chatbot app quickly rose to number one on the Apple App Store, causing excitement and some worries in the tech world.
Things got even more interesting when DeepSeek’s rise was linked to the drop in NVIDIA’s share price.
Well, the reasons for the fall can be many but its correlation indeed lit a wildfire of thrill and curiosity.
And when Sam Altman, the founder of OpenAI, even praised DeepSeek, especially for what they’re able to offer at such a low price. It grabbed the attention of many worldwide.
“A Chinese-made artificial intelligence (AI) model called DeepSeek has shot to the top of Apple Store's downloads, stunning investors and sinking some tech stocks.
Its latest version was released on 20 January, quickly impressing AI experts before it got the attention of the entire tech industry - and the world.
US President Donald Trump said it was a "wake-up call" for US companies who must focus on "competing to win".
What makes DeepSeek so special is the company's claim that it was built at a fraction of the cost of industry-leading models like OpenAI - because it uses fewer advanced chips.
That possibility caused chip-making giant Nvidia to shed almost $600bn (£482bn) of its market value on Monday - the biggest one-day loss in US history.”
— Reported by BBC
DeepSeek’s journey so far is unlike some other AI tools that are solely focused on the next marvel of AI or the debate between AI and humans.
It has deep ties to a Chinese hedge fund, High-Flyer Capital Management, which raises a series of censorship questions. Besides, its meteoric rise and aggressive pricing have opened war rooms and given AI and tech giants reasons to ponder.
In this article, I will share my personal experiences and how it is about to change the AI landscape.
Though Perplexy AI who has DeeSeek integrated for users who use its Pro plan claimed it’s uncensored.
From Quant Trading to AI Research
A Hedge Fund Origin
DeepSeek is backed by High-Flyer Capital Management, a Chinese quantitative hedge fund founded in 2015 by AI enthusiast Liang Wenfeng. Wenfeng, who started experimenting with trading algorithms as a student at Zhejiang University, officially launched High-Flyer as a hedge fund in 2019, focusing on AI-driven trading.
Spinning Off DeepSeek
In 2023, High-Flyer spun off a separate AI research lab—also called DeepSeek—to concentrate on advanced AI tools. Today, DeepSeek operates as its own company but still enjoys considerable support and funding from High-Flyer.
Young, Ambitious Technical Team
DeepSeek aggressively recruits PhD-level AI researchers from top Chinese universities. However, they also hire people with no computer science background to bolster the AI’s diverse “knowledge base.” This youth-driven, interdisciplinary approach has become a hallmark of DeepSeek’s rapid innovation.
Behind DeepSeek’s Rapid Rise
Hurdles with Hardware
While DeepSeek built its own data centers from the start, it hit a snag due to U.S. export bans on the most powerful chips.
As a result, the company resorted to using Nvidia H800 GPUs, a downscaled version of Nvidia’s H100 available to U.S. firms. Even so, DeepSeek found ways to keep training efficiently.
Compute-Efficient Techniques
Despite this hardware limitation, DeepSeek’s research has centered on methods that use far less computing power than rivals—an approach that keeps costs low. This is one reason Wall Street analysts and technologists are questioning whether the U.S. can maintain its AI lead and whether heavy demand for AI chips will continue.
The Models: Coder, LLM, Chat, and More
The Early Days (DeepSeek Coder, DeepSeek LLM, and Chat)
DeepSeek introduced its initial AI models in November 2023—DeepSeek Coder, DeepSeek LLM, and DeepSeek Chat. These models were respectable but didn’t cause big waves.
The Breakthrough (DeepSeek-V2)
Then came DeepSeek-V2 in spring 2024, a general-purpose text and image model that quickly grabbed attention for its high performance and incredibly low running costs.
DeepSeek’s domestic competitors—like ByteDance and Alibaba—were forced to drastically cut their own AI usage fees in response.
DeepSeek-V3: Another Leap
Released in December 2024, DeepSeek-V3 built on V2’s success, outperforming many established “closed” models (like OpenAI’s GPT-4o) and open-source alternatives (like Meta’s Llama). Its efficiency remained a key selling point.
DeepSeek R1: A “Reasoning” Powerhouse
In January 2025, DeepSeek unveiled R1, a specialized reasoning model on par with OpenAI’s GPT-o1 in tasks requiring deeper thought.
By design, R1 effectively “fact-checks” itself, taking a few extra seconds or minutes to arrive at a conclusion but offering more reliable results in physics, science, and math.
The trade-off is its slower response time—an acceptable compromise for many professionals needing accurate, step-by-step solutions.
DeepSeek's R1 model has demonstrated remarkable capabilities, outperforming industry leaders like ChatGPT while operating at a fraction of the cost.
The Chinese startup's AI chatbot app quickly rose to prominence, overtaking ChatGPT to become the top free app on Apple's App Store.
This sudden success has shattered the perception that smaller AI companies couldn't compete with tech giants, disproving OpenAI CEO Sam Altman's assertion that such competition was "hopeless".
DeepSeek's breakthrough highlights the potential for more cost-effective and efficient AI development, potentially accelerating widespread adoption of AI technologies.
The Controversy Around Censorship
Core Socialist Values
As I’ve mentioned earlier, as a Chinese-developed AI, DeepSeek’s models fall under China’s internet regulations. In DeepSeek’s publicly available chatbot, R1 refuses to answer politically sensitive questions (e.g., topics like Tiananmen Square).
Self-Hosted vs. Publicly Available
However, developers can download DeepSeek’s models under permissive licenses for commercial use, meaning the AI can be “uncensored” when self-hosted. That said, anyone using the company’s official app or API will still encounter the regulated (censored) version.
Business Model — or Lack Thereof?
Low or No Pricing
DeepSeek has rattled the market by either drastically undercutting the competition or offering its services for free. The company claims its efficiency breakthroughs allow them to maintain such extreme cost competitiveness.
Skepticism and Rapid Adoption
Some experts doubt the company’s cost figures. Regardless, developers have flocked to DeepSeek’s models.
On Hugging Face, there are already over 500 “derivative” models of R1 with a combined 2.5 million downloads.
API Pricing
For commercial users tapping the DeepSeek API (e.g., for coding tasks), the cost is $0.55 per million input tokens and $2.19 per million output tokens—a fraction of OpenAI’s $15 and $60 pricing tiers. For casual chatbot use in the official DeepSeek app or website, it remains free.
Going Head-to-Head with ChatGPT
DeepSeek vs. ChatGPT: Features
DeepSeek: Emphasizes a simple chatbot experience and a powerful search function. Lacks advanced features like image generation or “Canvas” tools. Offers two models in the chatbot:
DeepSeek-V3: General-purpose
R1 (DeepThink): Slower but better at math, science, and multi-step reasoning
ChatGPT: Offers more bells and whistles (e.g., image uploads with advanced analysis, subscription plans with multiple model choices, and integration with third-party tools).
Search Capabilities
Both DeepSeek and ChatGPT now support a conversational search function. DeepSeek’s search is built into the chatbot—just click or tap “Search” and continue the conversation. ChatGPT organizes its search results with more clarity and provides a sidebar for citations. DeepSeek’s approach is simpler and slightly less organized.
Who Should Switch?
If your needs revolve around everyday prompts, coding, or fact-finding—and you don’t mind missing advanced add-ons—DeepSeek’s free chatbot might suffice. But if you rely on uploading images for analysis or want specialized GPT tools, ChatGPT remains ahead.
Privacy and Security Concerns
Data Privacy Worries
Given its Chinese ownership, some security experts and the general public have raised concerns that user data could be exposed to the Chinese government. DeepSeek’s stance: user data is secure, and they use data centers in the United States and Europe for model hosting whenever possible outside China.
Malicious Attacks and Outages
DeepSeek has experienced malicious attacks leading to service interruptions. To manage demand, they’ve restricted who can sign up at times. Users eager to test the free chatbot may occasionally see “high demand” error messages.
What are the potential benefits of DeepSeek's model for smaller tech companies
DeepSeek's R1 model offers transformative benefits for smaller tech companies:
Cost Efficiency
Enterprises can now experiment and build AI prototypes more affordably, with the expectation of eventually scaling their applications at reduced costs. Specifically, DeepSeek's pricing is dramatically lower - just $2.19 per million output tokens compared to OpenAI's $60, and its capabilities are sufficient for most enterprise applications.
Democratized Access
DeepSeek offers its model and code freely, which:
- Promotes open-source collaboration
- Accelerates AI progression
- Makes AI development accessible to smaller companies and resource-constrained nations
- Encourages a collaborative global AI ecosystem
Innovation Opportunities
Smaller businesses can now access advanced AI technologies previously out of their financial reach. This opens up space for:
- Teamwork across borders
- New ideas
- Developers and entrepreneurs from diverse backgrounds contribute to AI growth
Strategic Advantage
DeepSeek's model signifies a potential disruption in AI economics, creating a more leveled playing field where innovation is not limited to the wealthiest entities, and fostering a more inclusive technological environment.
Can deepseek create images?
Yes, DeepSeek has recently launched its Janus Pro image generator with impressive capabilities. The Chinese AI company unveiled a new image generator that claims to beat OpenAI's DALL-E 3 and Stability AI's Stable Diffusion in benchmark tests.
Key details about DeepSeek's image generation:
The Janus Pro models are available on GitHub and Hugging Face in one-billion and seven-billion parameter versions, with the seven-billion parameter model reportedly competing with Stable Diffusion and DALL-E 3.
Currently, image uploads are restricted to 384 x 384 pixels, though Hugging Face demos can produce 768 x 768 pixel images. Early examples suggest the Janus Pro 7B is a significant upgrade from the original Janus model.
The image generator is free-to-use without restrictions and was developed at a remarkably low cost - less than $6 million, compared to billions spent by other AI companies.
Unique advantages include unprecedented creative freedom, allowing users to transform text descriptions into stunning, unique visuals across various styles - from fantastical landscapes to photorealistic portraits.
What makes DeepSeek's Janus-Pro 7B better than DALL-E 3 and Stable Diffusion
DeepSeek's Janus-Pro-7B demonstrates impressive capabilities across several key dimensions:
Benchmark Performance
The model achieved:
80% overall accuracy in text-to-image tasks
99% single-object accuracy
90% positional alignment
Compared to:
DALL-E 3: 67% accuracy
Stable Diffusion: 74% accuracy
Unique Technical Advantages
Janus-Pro-7B is a flexible "novel autoregressive framework" that can:
- Process image inputs
- Generate new images
- Combine capabilities of task-specific models in a single framework
This open-source approach is particularly noteworthy.
Training Methodology
The researchers:
- Introduced a decoupled architecture for visual understanding
- Leveraged synthetic data for enhanced training
- Achieved state-of-the-art performance through advanced data scaling techniques
Comparative Image Quality
While the model has nuanced performance:
Janus-Pro-7B images often look more realistic
It struggles with generating human figures
The model potentially produces images that more closely resemble real-world scenes compared to DALL-E 3.
Limitations
However, it's crucial to note that while the model shows high prompt adherence, this doesn't necessarily translate to aesthetic quality. Some critics argue that many of its generated images are subpar.
Integration with Perplexity AI—and the Backlash
Perplexity AI Partnership
San Francisco-based Perplexity AI has integrated DeepSeek R1 for its Pro users, offering yet another popular AI model option alongside OpenAI and Anthropic’s Claude.
Censorship Concerns
Some users criticized Perplexity AI, assuming the Chinese version of DeepSeek means censorship or data funneling to the Chinese government. Perplexity’s CEO, Aravind Srinivas, clarified on LinkedIn and X (formerly Twitter):
“All DeepSeek usage in Perplexity is through models hosted in data centers in the USA and Europe. DeepSeek is open-source. None of your data goes to China.”
Uncensoring the Model
Although DeepSeek’s publicly released models can refuse certain political topics, a self-hosted environment allows developers to remove these restrictions. Screenshots posted by Srinivas show DeepSeek R1 running uncensored within Perplexity’s infrastructure.
Meta Mobilises ‘War Rooms’ to Counter DeepSeek’s AI Breakthrough
Meta has created four specialized teams to tackle the competitive challenge posed by DeepSeek's AI progress. These teams are focused on:
- Analyzing DeepSeek’s strategies for reducing AI development costs
- Investigating the sources of its training data
- Exploring possible updates to Llama’s architecture
- Improving Meta’s AI products, including Llama 4
This step shows Meta's dedication to AI innovation, cost-effectiveness, and staying ahead of emerging competitors in the fast-evolving AI landscape.
Looking Ahead
DeepSeek’s rapid climb—culminating in its chatbot hitting number one on the iPhone App Store—has demonstrated just how swiftly AI upstarts can reshape the market. While its next generation of models is almost inevitable, questions remain:
- Will U.S. chip export restrictions tighten further?
- How will DeepSeek address global censorship concerns?
- Can DeepSeek’s ultra-low-cost strategy sustain itself?
What is clear is that DeepSeek, backed by a resourceful hedge fund and a young, ambitious engineering team, has introduced fierce new competition to entrenched AI giants. Whether DeepSeek can maintain its momentum under growing scrutiny is a story unfolding in real time—one that promises to shape the conversation around AI, regulation, and data security in the months and years to come.
