January 27, 2025

DeepSeek-R1: China's response to OpenAI

The Chinese AI startup DeepSeek has achieved a milestone in the Open-Source model DeepSeek-R1 sector of the AI industry. This model impresses with outstanding performance and a cost-efficient structure, making it a strong competitor. It provides alternatives to leading proprietary models such as OpenAI's o1 and stands out due to its efficiency.

DeepSeek-R1: What makes it special?

On January 20, 2025, DeepSeek introduced its new model: DeepSeek-R1, an Open-Source reasoning model with 671 billion parameters. This model stands out due to its exceptional performance and easy accessibility, drawing significant attention. Key features include impressive efficiency, high scalability, and wide applicability across various fields.

Outstanding performance

DeepSeek-R1 competes against OpenAI's o1 model in numerous areas and outperforms it in some cases. The results are clearly convincing:

79.8% Pass@1 on the AIME-2024 benchmark.
97.3% Score on the MATH-500 test.
An Elo rating of 2,029 on Codeforces, surpassing 96.3% of human participants.

These benchmarks prove that DeepSeek-R1 is powerful and versatile, from math and logic to programming competitions.

Cost efficiency

DeepSeek offers its API up to 95% cheaper than OpenAI and impresses with attractive prices. While OpenAI’s o1 model costs $7.50 per million input tokens, DeepSeek-R1 charges only $0.14. Additionally, DeepSeek provides the model on its web and app platforms free of charge.

Open source nature

DeepSeek-R1 is available under an MIT license on Hugging Face and is freely accessible to everyone. The model can be used for commercial applications and extensively researched by scientists, creating numerous opportunities. It stands out clearly from the closed models of large providers like OpenAI and promotes innovation. The community can further develop the model and experiment with new approaches, which additionally drives progress.

How DeepSeek has learned from OpenAI

Interestingly, there are indications that DeepSeek was inspired by OpenAI technologies. It may even have used them directly.

Behavioral Similarity: Earlier DeepSeek models, such as DeepSeek V3, occasionally identified themselves as "ChatGPT" or "GPT-4." This suggests a behavioral similarity.
Training on OpenAI Outputs: Experts suspect that DeepSeek may have trained its models using OpenAI outputs. This process is known as "Knowledge Distillation." At the same time, this approach saves significant costs in model development.

These findings show how start-ups can benefit from existing technologies. This enables them to create powerful alternatives.

Impact on the AI industry

DeepSeek-R1 initiates an important shift in global AI competition and offers exciting opportunities for innovation. It creates new possibilities for competition and advancement that can sustainably shape the field of AI.

Democratization of AI

With open access and low costs, DeepSeek-R1 enables smaller companies and developers to access advanced AI technology. This lowers the barriers to entry into the AI industry and promotes more innovation and competition.

Cost reduction

DeepSeek developed DeepSeek-V3 in just two months, using a budget of only $5.58 million. This represents only a fraction of the resources typically required by US competitors like OpenAI. This demonstrates that powerful AI models do not necessarily require billions in investments to be successfully developed.

Global competitiveness

DeepSeek highlights China's increasing strength in AI development and demonstrates its growing influence in this field. This could accelerate the dynamics in the "AI arms race" between China, the USA, and other leading AI nations.

Open-source innovation

The openness of DeepSeek-R1 gives researchers and developers the opportunity to further develop and adapt the model. In contrast, closed systems such as OpenAI's o1 only offer very limited access to the underlying models.

Challenges and controversies

Originality concerns

The suggestion that DeepSeek may have used existing OpenAI models raises important ethical questions. Although “distillation” is a common practice, legal conflicts could arise if the use of protected OpenAI content is proven.

Data protection and security

As an open source model, DeepSeek-R1 comes with potential risks, such as misuse to create false information or malicious applications.

Sustainability of the business model

Providing the model free of charge could raise questions about the profitability of DeepSeek in the long term. The sustainability of such a model depends on whether more profitable monetization strategies can be established.

Conclusion: A turning point for the AI industry

With DeepSeek-R1, the Chinese start-up shows that AI innovation is no longer exclusively dominated by large tech companies such as OpenAI or Google. Open source models such as R1 are creating new opportunities for smaller companies, researchers and developers who want to benefit from low-cost access and great flexibility.

Such developments could further diversify the future of AI and boost both the speed of innovation and competition in the industry. DeepSeek proves that impressive models are possible even with limited resources, which sends an important message to the global AI market.