Artificial Intelligence, Privacy

The ‘DeepSeek Moment’ and what it means for the healthcare industry

Igor Alvarez - Chief Innovation Officer

DeepSeek logo applied on a cellphone screen.
a hand touching a computer screen full of colorful zeoes and ones. Concept image for advanced technology.

Recently, DeepSeek has been making headlines worldwide, extending beyond tech news into mainstream media. Beyond its significant market impact on NVIDIA and U.S. tech giants, what does this mean for the rest of us, particularly in the healthcare sector? To understand DeepSeek’s potential impact on our sector, let’s examine what it is, evaluate the strengths and limitations of its recently released model r1, and determine how—or if—we should use it in our daily projects, considering existing similar tools in the market.

What is DeepSeek and what is r1?

DeepSeek is a Chinese company founded in 2023 in Hangzhou that focuses on AI and Large Language Model (LLM) development. In November 2023, they released their first product: DeepSeek Coder, a family of open-source LLM models designed to generate code. Trained on more than 80 programming languages, it achieved State-of-the-Art (SOTA) performance compared to similar models. Since then, DeepSeek has consistently released open-source models that achieved SOTA status in their respective categories. However, these models weren’t without limitations. For instance, their DeepSeek LLM, released in November 2023, could compete with the then-SOTA model GPT-4 but faced performance and scalability challenges.

After numerous releases from various players in the LLM market (including some from DeepSeek itself), OpenAI launched their new model, o1, in December 2024. This model promised to revolutionize how LLMs and AI would work by “thinking through” each task methodically, questioning its approach at every step before delivering a final answer. This innovative approach combined Chain-of-Thought (CoT) reasoning with test-time computing. Unlike traditional models that used most computing resources during training, this new approach would also leverage computational power during the answer phase, producing superior results at the cost of increased resource requirements.

Despite an initial slow and restricted rollout, o1’s results were superior to predecessors like Claude Sonnet 3 and GPT 4o. Particularly in logic, math, and code-related tasks, the model surpassed all previous models and achieved exceptional performance across various benchmarks. This success led many skeptics, who had been predicting a new ‘AI Winter’, to reconsider their position. However, the price was substantial: accessing o1’s ‘pro’ version (with limited daily use) required a $200/month OpenAI subscription. A more affordable version, suitable for most use cases, was available for $20, yet with usage limitations.

How is DeepSeek r1 different?

DeepSeek r1, released on January 20, 2024, promised results comparable to OpenAI’s o1 at a fraction of the cost. The model is currently free to use —even commercially— and has open weights, allowing anyone to download the models and run it locally. Released under an MIT license, it came with a free web-based chat interface, iOS and Android apps, and a family of smaller models. These smaller versions, trained on the main model, can even run on standard consumer hardware at home, without any additional cost.

Even the developer-oriented API access costs significantly less than o1

DeepSeek API access price comparison with OpenAI o1, the endpoint used by programmers, costs only a fraction of OpenAI o1: Input: $0.55/1M Tokens vs. $15 for OpenAI o1 and Output: $2.19/1M Tokens vs. $60 for OpenAI o1.

DeepSeek API access, the endpoint used by programmers, costs only a fraction of OpenAI o1:
Input: $0.55/1M Tokens vs. $15 for OpenAI o1 and Output: $2.19/1M Tokens vs. $60 for OpenAI o1.

DeepSeek API access, the endpoint used by programmers, costs only a fraction of OpenAI o1: Input: $0.55/1M Tokens vs. $15 for OpenAI o1 and Output: $2.19/1M Tokens vs. $60 for OpenAI o1.

One particularly noteworthy aspect, and maybe what caused the big shake in the financial market, is the cost efficiency of DeepSeek’s development. While OpenAI reportedly invested $500 million in ‘GPT-5’ (Project Orion) development and maintains estimated annual operating costs exceeding $3 billion, DeepSeek r1’s total training investment was only $5.6 million.

But how capable is r1 in terms of quality? In the AI industry, companies evaluate their models’ abilities through standardized Q&A tests called “benchmarks.” These tests, created by domain experts, provide a reliable way to compare models from different companies. Bellow we can see some of the benchmark results shared by the DeepSeek team:

DeepSeek r1 benchmark comparison graph with OpenAI o1: DeepSeek surpasses OpenAI o1 by a small margin in 3/6 of the mentioned benchmarks, and get very close results in the remaining ones.

DeepSeek r1 surpasses OpenAI o1 by a small margin in 3/6 of the mentioned benchmarks, and get very close results in the remaining ones.

According to these benchmarks, DeepSeek r1 matches or outperforms OpenAI o1 across all tests. Even r1 32B, one of the smaller models in the r1 family designed for consumer hardware, performs as well as or better than o1-mini. But what does this mean in practical terms? And does r1 have any limitations for your daily use?

The case of r1 for Healthcare

The use of AI and LLMs in healthcare is well-established, with numerous proven applications in diagnosis, patient relations, and—as covered in our previous article—even drug discovery research. But let’s examine how r1 fits into this ecosystem and its specific considerations for healthcare applications.

Is there any privacy issue?

When discussing r1, privacy is often the first concern, given that DeepSeek is a Chinese company operating in China. While these concerns are valid in many contexts, r1 offers a unique privacy and security advantage that even OpenAI and Google’s flagship models previously lacked: as an open model, organizations can host it locally on servers within US/EU territories, or even within their own company if they have the resources to do so. This approach makes it easier to handling PII (Personal Identifiable Information) and medical data ensuring complete privacy and enhanced security and may even be benefitial for companies looking to apply any LLM pipeline to secret/intern information. Unlike using standard LLMs through OpenAI, Claude, or Google’s APIs, local hosting guarantees that sensitive data EVER leaves your infrastructure or gets shared with third parties.

Is it easy to implement?

Despite its different ‘reasoning’ and Chain-of-Thought approach, DeepSeek r1 serves as a powerful tool for any application where you’d typically use a standard LLM. The transition to self-hosting and slightly different implementation requires minimal additional effort. Our recent tests in text classification, programming, and educational content generation tasks have shown r1’s exceptional performance, validating the official benchmarks—the model proves more reliable, logical, and capable than its predecessors.

How can I add it to my daily tasks/products?

Like all LLM models, r1 may produce hallucinations when its knowledge is limited, and its answers cannot be guaranteed 100% accurate. While the model’s capabilities are impressive and its Chain-of-Thought process offers transparency, implementing additional security measures remains crucial—as with any LLM-based tool.

At C/Edge, our team integrates LLM technology with RAGs (Retrieved Augmented Generations), Function Calling, Safety-Models, Agentic pipelines, and other techniques to address these challenges and optimize performance for your team or client. Our goal extends beyond providing a simple chatbot interface—we deliver a comprehensive solution grounded in your chosen real-world data that aligns with your brand’s expectations.

Are you interested in learning more?

I will be conducting a series of webinars this year as C/Edge’s new Director of Innovation. In these sessions, I’ll share my expertise on technologies and new developments in the digital landscape for the healthcare communication and advertising market. Stay tuned for more information on the dates and topics of the webinars!

Conclusion

DeepSeek r1 represents a significant milestone in AI development, not just for its technical achievements but for what it means for the healthcare industry’s future. By combining state-of-the-art performance with unprecedented accessibility—both in terms of cost and deployment options—r1 has the potential to democratize advanced AI capabilities across the healthcare sector. While the model’s arrival has sparked discussions about everything from market dynamics to privacy concerns, its practical benefits for healthcare organizations are clear: the ability to leverage cutting-edge AI technology without compromising on data security or breaking the budget. As we continue to explore and implement these technologies at C/Edge, our focus remains on turning these technological advancements into practical, secure, and effective solutions for our healthcare clients. The emergence of models like r1 suggests that the future of healthcare AI will be defined not just by capability, but by accessibility and adaptability to real-world healthcare needs.