DeepSeek: The Complete Story of Its Birth and Development

Powered By: Charlie
Date: February 1, 2025

Origins

DeepSeek emerged from the futuristic vision and technical expertise of Liang Wenfeng, an entrepreneur born in Guangdong and educated in Zhejiang. Before founding DeepSeek, Liang had already demonstrated his skills in mathematics and engineering, which led him to co-found the hedge fund High-Flyer Quant in 2015. This fund, specializing in using machine learning models to operate in financial markets, allowed him to gain experience and, most importantly, resources in Nvidia GPUs.

"In 2021, Liang began purchasing thousands of GPUs, acquiring up to 10,000 units, with the goal of exploring the potential of artificial intelligence."

These acquisitions and the experience gained at High-Flyer laid the foundation for the ambitious project that would later become DeepSeek.

Founding of DeepSeek

In 2023, leveraging the know-how and infrastructure developed at High-Flyer, Liang Wenfeng founded DeepSeek. The core idea was to develop large language models (LLMs) that could compete with tech giants like OpenAI and Google but at a fraction of the cost. DeepSeek is built on three fundamental pillars:

1. Cost efficiency: While training comparable models required investments of hundreds of millions of dollars, DeepSeek managed to develop its DeepSeek-R1 model with approximately $5.6 million.
2. Open-source approach: Unlike closed models, DeepSeek decided to openly share its code and documentation, fostering collaboration and transparency.
3. Resource optimization: With the massive stock of GPUs acquired at High-Flyer and innovative training techniques, DeepSeek overcame limitations imposed by chip export restrictions.

"DeepSeek was conceived as a pure research lab, where the priority was not immediate profit but pushing the boundaries of AI knowledge."

Technical Development and Releases

Initial Models and Early Evolution

During the first months after its founding, DeepSeek’s team, mostly composed of young graduates from China’s elite universities, experimented with Transformer-based architectures. Resource utilization was optimized through techniques such as:

Context extension: Proprietary methods (such as YaRN) were used to increase the model’s capacity, enabling it to process very long text sequences.
Mixture of Experts (MoE) and Low-Rank Attention (MLA): These techniques allow only the necessary “experts” to be activated for each task, maximizing efficiency.
Mixed-precision training: The adoption of 8-bit and 12-bit arithmetic for certain operations significantly reduced computational costs.

The Groundbreaking Launch of DeepSeek-R1

In early 2025, DeepSeek launched its model DeepSeek-R1, specialized in complex reasoning tasks, mathematics, and code generation. This model positioned itself as a direct rival to OpenAI’s systems, triggering ripple effects in the tech market, including a historic drop in Nvidia’s stock price.

"The R1 model, developed with minimal investment, has been described as an 'AI Sputnik moment,' proving that reasoning capabilities comparable to tech giants can be achieved."

Key Innovations

DeepSeek implemented several technical innovations that set it apart:

Context length adjustment: Enables processing of up to 128k tokens, ideal for extensive documents.
Use of MoE and MLA techniques: Optimizes GPU usage by activating only the necessary modules for each task.
Mixed-precision training: Reduces memory and energy consumption without sacrificing performance.

Additionally, its commitment to open-source has allowed other researchers to replicate and improve its advancements, accelerating innovation in the field.

Impact and Future Prospects

Market Impact and Competition

DeepSeek’s emergence has challenged the paradigm of massive investment in AI hardware. Achieving high-performance results at reduced costs has forced investors and competitors to rethink their strategies. This movement has created a domino effect, even impacting semiconductor companies like Nvidia.

Geopolitical Implications

DeepSeek’s success is not just technical but also strategic. By circumventing U.S. export restrictions on chips, DeepSeek positions itself as a key player in the global AI race, potentially reshaping the technological balance between East and West.

Culture of Open Innovation

The commitment to open-source and collaboration is another major differentiator for DeepSeek. This strategy not only promotes transparency but also drives collaborative development and technology adaptation to different contexts and needs, potentially accelerating global AI advancements.

Conclusion

The story of DeepSeek is that of a visionary entrepreneur, Liang Wenfeng, who has successfully transformed the experience and resources accumulated at High-Flyer into a disruptive AI project. From its origins in 2015 to DeepSeek’s establishment in 2023 and the groundbreaking launch of its R1 model in 2025, the company has proven that AI innovation is not solely dependent on massive investments but also on creativity, efficiency, and openness.

DeepSeek not only offers a competitive alternative to traditional models like GPT-4 but also introduces a new development paradigm based on efficiency, open-source, and global collaboration. This approach could define the future of AI, driving significant changes in both the market and AI geopolitics.

Sources

Articles and reports from El País, The Guardian, HuffPost, and Wikipedia
Technical documents and publications from DeepSeek on GitHub and Hugging Face

This narrative reflects the convergence of ingenuity, strategy, and boldness in building one of the most disruptive AI startups today.