DeepSeek: Pioneering China’s AI Revolution

Dr.N.Sriram,

Founder & Managing Director, Genesis Research Labs, S.R. Nagar, Hyderabad, Telangana, India.

Contact: +91 9704361947

Email: drnsriram@gmail.com

DeepSeek is a Chinese artificial intelligence lab specializing in open-source large language models (LLMs). Founded and backed by the hedge fund High-Flyer, DeepSeek has rapidly emerged as a major player in the AI landscape, with its headquarters based in Hangzhou, Zhejiang. Here’s a closer look at the journey of DeepSeek and its transformative contributions to AI development.

The Foundation of DeepSeek

The story begins in February 2016, when Liang Wenfeng, a Zhejiang University alumnus and AI enthusiast, co-founded High-Flyer. Liang, who began trading during the 2007-2008 financial crisis, developed a keen interest in AI-driven financial systems. By 2019, High-Flyer pivoted to focus on AI trading algorithms exclusively, and by 2021, all trading operations were fully automated with AI.

In April 2023, High-Flyer expanded its ambitions beyond finance, establishing an Artificial General Intelligence (AGI) research lab. By May 2023, this lab was spun off as an independent entity, DeepSeek, supported primarily by High-Flyer and Liang’s vision. While venture capital firms were hesitant to fund the venture due to its long-term focus, DeepSeek persevered, marking its entry into the competitive AI arena.

A Catalyst for China’s AI Price War

In May 2024, DeepSeek launched DeepSeek-V2, an LLM that outperformed competitors in cost efficiency. Dubbed the “Pinduoduo of AI” for its aggressive pricing strategy, DeepSeek forced major Chinese tech giants like ByteDance, Tencent, Baidu, and Alibaba to reduce their AI model prices. Despite its lower pricing, DeepSeek’s models were profitable, a stark contrast to its rivals.

DeepSeek’s focus remains firmly on research and development, with no immediate plans for commercialization. The company has also adopted an unconventional hiring strategy, prioritizing technical skills over experience, resulting in a workforce primarily composed of recent graduates and early-career developers.

Major Milestones in DeepSeek’s Evolution

DeepSeek LLM

• November 2, 2023: The release of DeepSeek Coder, its first open-source model, available for free to researchers and commercial users under the MIT license.
• November 29, 2023: DeepSeek LLM, a 67-billion-parameter model, debuted alongside DeepSeek Chat, a chatbot application. While it achieved performance comparable to GPT-4, scalability and efficiency challenges persisted.

V2: A Turning Point

• May 2024: DeepSeek-V2 emerged as a cost-effective and competitive model, pricing at just 2 RMB per million output tokens. It ranked seventh on the University of Waterloo Tiger Lab’s LLM leaderboard, cementing DeepSeek’s reputation as an AI disruptor.

V3: A Technological Leap

• December 2024: DeepSeek introduced DeepSeek-V3, a 671-billion-parameter model trained on 14.8 trillion tokens. Despite its vast scale, training costs were remarkably low at $5.58 million, showcasing DeepSeek’s resource optimization. The model rivaled industry leaders like GPT-4 and Claude 3.5 Sonnet while surpassing others like Llama 3.1.

R1: Advancing Specialized Reasoning

• November 2024: The R1-Lite-Preview was released, targeting mathematical reasoning and logical inference. Despite outperforming competitors on benchmarks like AIME, real-world testing showed mixed results.
• January 20, 2025: The release of DeepSeek-R1 and R1-Zero introduced models tailored for advanced reasoning tasks. R1-Zero, trained solely with reinforcement learning, demonstrated innovative training methodologies like Group Relative Policy Optimization (GRPO). However, challenges in language consistency and readability persisted.

Impact and Challenges

DeepSeek’s success highlights the growing capabilities of China’s AI ecosystem, even amidst geopolitical constraints like U.S. sanctions. The launch of V3 and R1 has been described as a “Sputnik moment” for American AI, underscoring the rapid advancements in China’s AI industry.

While DeepSeek’s journey is still unfolding, its achievements in cost efficiency, open-source accessibility, and large-scale AI development have already left a significant mark. As the company continues to refine its models and explore new horizons, its role in shaping the global AI landscape will undoubtedly expand.