On January 27, 2025, DeepSeek-R1, a new Generative AI model, was introduced and rapidly gained adoption, taking over the user market to become the No.1 downloaded AI app [1], [2], [3]. DeepSeek is notable because it is open-sourced under the MIT license, is cost-effective, and exhibits impressive reasoning skills, thanks to techniques like Mixture of Experts, reinforcement learning, FP8 mixed precision framework, and multi-stage training, among others [4]. Further, DeepSeek-R1 outperforms GPT-4o and OpenAI o1 across various benchmarks and excels in tasks such as mathematics and coding [5], [6]. More importantly, this development highlights the trend that “open source models are surpassing proprietary ones” [7].

The emergence of DeepSeek has ignited extensive discussions. Much has been written about what sets DeepSeek apart and why it could help write the blueprint for the next era of AI [7], [8], [9]. In this paper, we consider how DeepSeek, along with its underlying new technology and framework, could define the blueprint for the next era of AI in healthcare. We will focus on its performance, reasoning capabilities, training methods, usage, and ethical considerations.