DeepSeek’s success with open-source AI

Key Points

• DeepSeek, a Chinese AI company founded in July 2023, has made significant strides in AI, particularly with its model DeepSeek-R1, released in January 2025.
• It seems likely that DeepSeek’s success stems from its cost-efficient training methods, using reinforcement learning and a mixture-of-experts architecture, costing around US$6 million compared to US$100 million for similar models.
• Research suggests DeepSeek’s open-source approach, under the MIT License, has boosted community engagement and user base, with its AI assistant app topping Apple’s App Store.
• The evidence leans toward DeepSeek impacting the global AI landscape by challenging US tech giants, causing stock market reactions, and promoting accessible AI development.
• There is uncertainty around DeepSeek’s long-term plans and ethical AI practices, but it appears focused on expanding and improving models.

Company Background and Rise

DeepSeek was established in Hangzhou, China, in July 2023 by Liang Wenfred, also co-founder of the hedge fund High-Flyer, which fully funds and owns DeepSeek. This unique funding model allowed DeepSeek to prioritize long-term AI research without external investor pressure. Starting as a hedge fund initiative, DeepSeek has risen to become a notable AI leader, challenging established players with its efficient and innovative models.

Path to Developing R1 and Training Innovations

DeepSeek’s path to developing DeepSeek-R1 involved a multi-stage process. It began with DeepSeek-R1-Zero, using pure reinforcement learning without supervised fine-tuning, achieving reasoning capabilities but with readability issues. DeepSeek-R1 refined this by adding supervised fine-tuning on curated data, followed by further reinforcement learning, resulting in a model competitive with OpenAI’s GPT-4o and o1 at a fraction of the cost. Their innovative training methods, including a mixture-of-experts architecture, reduced computational costs, with DeepSeek-V3 trained for about US$6 million compared to US$100 million for OpenAI’s GPT-4 (DeepSeek – Wikipedia).

Success with Open-Source AI and Community Engagement

DeepSeek’s commitment to open-source AI, releasing models under the MIT License, has fostered significant community engagement. This approach allows free commercial and academic use, attracting researchers and developers. Its AI assistant app, powered by DeepSeek-R1, became the top-downloaded free app on Apple’s App Store in January 2025, highlighting strong user base growth (DeepSeek – Wikipedia).

Impact and Recognition

DeepSeek’s breakthrough in AI efficiency has disrupted the global AI landscape, challenging the dominance of US companies like OpenAI and Nvidia. Its low-cost models led to stock market reactions, with shares of tech giants dropping as investors reassessed AI valuations. Media and industry recognition have been substantial, with coverage in outlets like BBC and Forbes, positioning DeepSeek as a key player (What is DeepSeek – and why is everyone talking about it? – BBC News, All About DeepSeek — The Chinese AI Startup Challenging US Big Tech – Forbes).

Survey Note: DeepSeek’s Success Story – A Detailed Analysis

DeepSeek, a Chinese artificial intelligence company based in Hangzhou, Zhejiang, has emerged as a significant player in the global AI landscape since its founding in July 2023 by Liang Wenfred, who also co-founded the hedge fund High-Flyer. This survey note provides a comprehensive examination of DeepSeek’s journey, covering its path to developing R1, success with open-source AI, breakthrough in AI efficiency, impact on the global AI landscape, rise from hedge fund to AI leader, founder’s vision, early days and initial struggles, key partnerships, innovative training methods, contributions to AI research, future plans, competitive advantages, user base and community engagement, media and industry recognition, approach to ethical AI development, expansion into new markets, philanthropy and social responsibility, team and culture, customer success stories, and blog and thought leadership.

Company Background and Rise from Hedge Fund to AI Leader

DeepSeek operates as an independent AI research lab under the umbrella of High-Flyer, which provides sole funding. This model has enabled DeepSeek to pursue ambitious AI projects without the pressure of external investors, focusing on long-term research and development. The company’s rise from a hedge fund initiative to an AI leader is marked by its rapid development of competitive large language models (LLMs), challenging the dominance of US tech giants (DeepSeek – Wikipedia).

Path to Developing R1 and Innovative Training Methods

DeepSeek’s path to developing DeepSeek-R1, released in January 2025, involved a multi-stage training process. It started with DeepSeek-R1-Zero, using pure reinforcement learning (RL) without supervised fine-tuning (SFT), demonstrating reasoning capabilities such as self-verification and long chain-of-thought (CoT) reasoning. However, readability issues prompted the development of DeepSeek-R1, which incorporated SFT on curated “cold-start data” followed by additional RL stages. This approach allowed the model to achieve performance comparable to OpenAI’s GPT-4o and o1 at a fraction of the cost (GitHub – deepseek-ai/DeepSeek-R1, DeepSeek-R1: Features, o1 Comparison, Distilled Models & More | DataCamp).

DeepSeek’s innovative training methods include the use of a mixture-of-experts (MoE) architecture, where only a portion of the model’s parameters (e.g., 37B out of 671B for DeepSeek-V3) is activated per input, significantly reducing computational costs. This efficiency is evident in claims that training DeepSeek-V3 cost approximately US$6 million, compared to US$100 million for OpenAI’s GPT-4 in 2023, and using about one-tenth of the computing power for Meta’s Llama 3.1 (DeepSeek – Wikipedia, DeepSeek-V3: Efficient and Scalable AI with Mixture-of-Experts | Medium).

Success with Open-Source AI and Community Engagement

DeepSeek’s success with open-source AI is rooted in its release of models under the MIT License, enabling free commercial and academic use. This approach has fostered significant community engagement, with models like DeepSeek-R1 and its distilled versions (e.g., DeepSeek-R1-Distill-Qwen-32B) available for researchers to build upon. The company’s AI assistant app, powered by DeepSeek-R1, became the most-downloaded free app on Apple’s App Store in January 2025, reflecting a growing user base (DeepSeek – Wikipedia, What is DeepSeek, the Chinese AI startup that shook the tech world? | CNN Business).

Breakthrough in AI Efficiency and Competitive Advantages

DeepSeek’s breakthrough in AI efficiency is a cornerstone of its success, with training costs significantly lower than competitors. This efficiency is driven by its innovative training methods and MoE architecture, positioning DeepSeek with competitive advantages in cost and performance. The company’s ability to develop high-performing models with fewer resources challenges the notion that only large investments can yield top-tier AI, impacting industry perceptions (DeepSeek explained: Everything you need to know – TechTarget).

Impact on the Global AI Landscape and Media Recognition

DeepSeek’s impact on the global AI landscape is profound, challenging US tech dominance and causing market reactions. Its low-cost models led to stock market sell-offs for companies like Nvidia and Meta in January 2025, as investors reassessed AI valuations. Media and industry recognition have been extensive, with coverage in BBC, Reuters, and Forbes, highlighting DeepSeek’s disruptive potential (What is DeepSeek – and why is everyone talking about it? – BBC News, What is DeepSeek and why is it disrupting the AI sector? | Reuters).

Founder’s Vision and Early Days

Liang Wenfred’s – Founder of Deepseek AI

Liang Wenfred’s vision for DeepSeek is to achieve artificial general intelligence (AGI) with a focus on curiosity and long-termism, as stated on their website (DeepSeek – DeepSeek.com). In its early days, DeepSeek faced challenges in building a team and acquiring resources, but its focus on innovation and efficiency allowed rapid progress, with the first model released in November 2023 (DeepSeek explained: Everything you need to know – TechTarget).

Key Partnerships and Contributions to AI Research

While specific partnerships are not widely detailed, DeepSeek likely collaborated with hardware providers and research institutions to access computational resources. Its contributions to AI research are significant, with open-source models and research papers detailing methods like RL for reasoning and distillation techniques, benefiting the broader AI community (DeepSeek-R1 Release | DeepSeek API Docs).

Future Plans, Expansion, and Ethical Considerations

DeepSeek’s future plans include improving model performance, expanding capabilities, and exploring new applications, with a focus on open-source principles. Expansion into new markets is facilitated by its API and mobile app, making AI accessible globally. Its approach to ethical AI development is less documented, but models are designed to avoid politically sensitive topics, aligning with Chinese censorship standards, which may raise international ethical questions (What is DeepSeek, and why is it causing Nvidia and other stocks to slump? – CBS News).

Philanthropy, Team, and Culture

As a new company, DeepSeek’s philanthropy and social responsibility initiatives are not prominent, though its open-source contributions can be seen as community-oriented. The team comprises young, talented graduates from top Chinese universities, with a culture emphasizing technical ability over traditional experience, fostering innovation (All About DeepSeek — The Chinese AI Startup Challenging US Big Tech – Forbes).

Customer Success Stories and Thought Leadership

Customer success stories are emerging, with early users reporting positive experiences with the AI assistant app. DeepSeek maintains thought leadership through its website and social media, sharing research findings and updates, reinforcing its position in AI discourse (DeepSeek – DeepSeek.com).

Summary Table: Key Milestones and Metrics

This detailed analysis underscores DeepSeek’s rapid rise and its potential to reshape the AI industry, with a focus on efficiency, accessibility, and innovation.