DeepSeek: a low-cost, open-source challenger to industry giants

Unlike OpenAI’s proprietary AI models, DeepSeek-R1 is open source and free to use. This approach disrupts the existing monetization strategies of Silicon Valley tech companies, such as OpenAI, which charge users for access to AI-powered services like ChatGPT. With DeepSeek’s AI model available at no cost, organizations and developers can integrate powerful AI capabilities without being tied to costly subscription plans or usage-based fees. As a Chinese startup backed by Beijing’s growing AI ecosystem, DeepSeek’s move introduces a politically significant shift in the AI landscape, challenging the dominance of U.S.-based firms. This development could reshape the competitive field, making high-quality AI more accessible while forcing established players to rethink their revenue streams.

Reinforcement learning with chain of thought

DeepSeek-R1’s core innovation lies in its reinforcement learning (RL) approach. This method emphasizes reasoning tasks, allowing the model to refine its capabilities iteratively without supervised data. The researchers also employed a rule-based reward engineering system, which outperforms traditional neural reward mechanisms by guiding the model’s learning process more effectively. This setup ensures that DeepSeek evolves with minimal human intervention, unlocking its potential for self-improvement.

Multi-stage training and cold-start data

The development process includes a multi-stage training pipeline, beginning with cold-start data before transitioning to RL. This phased approach mitigates issues like language mixing and poor readability, which have been noted as challenges in earlier iterations. By incorporating these refinements, DeepSeek aims to enhance the reasoning accuracy and linguistic coherence of its models.

Distillation for efficiency

To further optimize performance, DeepSeek has employed model distillation techniques. This process compresses the model’s capabilities into a significantly smaller architecture, achieving high performance with only 1.5 billion parameters. Such innovations make the model both cost-effective and computationally efficient, broadening its appeal to a wider range of users.

One of DeepSeek’s standout features is its use of a mixture of experts (MoE) architecture. Unlike traditional dense models that activate all neurons during processing, MoE selectively activates specific components of the model depending on the task. This selective activation theoretically enhances speed and reduces computational overhead. Benchmarks, however, show mixed results. While the architecture is designed for efficiency, some users have reported server issues and slow response times, raising questions about its reliability under high-demand scenarios.

DeepSeek is emerging as a serious challenger to OpenAI, particularly in cost and accessibility. While OpenAI has built its reputation on high-performance, proprietary models like GPT-4, DeepSeek is focusing on efficiency and affordability. DeepSeek trained its latest model with remarkable efficiency, using just 2,000 GPUs over 55 days at a cost of $5.58 million—significantly less than OpenAI’s much higher training expenses. This resource efficiency extends to their API offerings, where they provide two primary models: DeepSeek-V3 (deepseek-chat), a general-purpose model trained on 15 trillion tokens, and DeepSeek-R1 (deepseek-reasoner), optimized for advanced reasoning, math, and coding tasks. As a result, DeepSeek’s API services are reportedly 20 to 40 times cheaper than OpenAI’s, making it an appealing choice for businesses aiming to reduce costs without sacrificing performance.

One other key contrast is OpenAI’s closed-source approach versus DeepSeek’s open-source strategy. OpenAI keeps its most powerful models proprietary, emphasizing safety, control, and enterprise partnerships, while DeepSeek has made its models publicly available, encouraging wider adoption and customization. Performance-wise, DeepSeek-R1 competes well in tasks like coding and mathematical reasoning, but OpenAI still has an edge in multi-modal AI, enterprise integrations, and long-term reliability.

Sam Altman has acknowledged DeepSeek’s progress, calling its work “impressive” but maintaining that OpenAI will continue improving its models. This competition could push OpenAI to accelerate its innovations, but for now, its track record and corporate trust keep it ahead in enterprise adoption.

Benchmark results show that DeepSeek-R1 performs exceptionally well across a range of tasks, often outshining other leading models like GPT-4o, OpenAI o1, and Claude-3.5. In English tasks, it achieves top scores in areas like MMLU-Redux (92.9 EM) and AlpacaEval2.0 (87.6 LC-winrate), showcasing strong contextual understanding and evaluation capabilities.

For coding, DeepSeek-R1 stands out with a high percentile on Codeforces (96.3) and strong results in LiveCodeBench (65.9 Pass@1-COT), demonstrating its ability to handle complex code generation and problem-solving tasks. In math benchmarks, it leads in MATH-500 (97.3 Pass@1) and performs impressively on AIME 2024 (79.8 Pass@1). Its performance in Chinese-language tasks, such as CLUEWSC (92.8 EM), highlights its effectiveness in multilingual scenarios.

These results demonstrate that DeepSeek-R1 is not just competitive but often superior in its precision, reasoning, and adaptability across diverse tasks.

For businesses, data privacy is a critical factor when selecting an AI service. DeepSeek’s storage policies and data handling practices are particularly relevant, given that data protection laws vary across jurisdictions.DeepSeek states that it follows recognized security standards, but its data storage location in China could be a concern for companies with strict compliance requirements. A 2024 cybercrime index also lists China as one of the most frequently targeted countries for cyberattacks, which may increase the risk for sensitive information. 

Another important aspect is access control. Under Chinese data regulations, authorities can request access to information stored within the country. This raises potential risks for businesses handling sensitive or regulated data, as they must carefully evaluate whether DeepSeek’s infrastructure aligns with their compliance needs.

Looking ahead, DeepSeek plans to address its current limitations and expand its capabilities. Areas of focus include:

• Language mixing: enhancing the model’s ability to process multilingual queries without defaulting to English or Chinese.

• Prompt engineering: developing robust zero-shot settings to improve performance in response to minimal input.

• Software engineering tasks: incorporating techniques like rejection sampling and asynchronous evaluations to boost efficiency in coding-related applications.

DeepSeek is shaking up the AI world with its budget-friendly, open-source model, giving more people access to powerful AI while challenging the industry's biggest players. While challenges remain in terms of scalability, speed, and privacy, the company’s innovations—from reinforcement learning to model distillation—signal a promising future for cost-effective, reasoning-focused AI solutions.