Artificial intelligence (AI) typically refers to computational technologies that mimic the mechanisms of human intelligence [1-4], such as thinking, deep learning, adaptation, interaction, and understanding [5]. Alternatively, it can be defined as a computational program or machine performing a task that, if accomplished by a human, would be considered intelligent [6]. AI technologies are being extensively used in the health sector to tackle the myriads of challenges confronting it. We use the term “AI health” to describe the application of AI technologies in the health sector to influence health outcomes. The fundamental distinction between AI health and digital health lies in the fact that, in the former, decisions and actions directly related to health (at both the individual and population level) are made or assisted by AI (computer programs and machines) rather than being independently made by humans. In other words, the defining characteristic of AI health is that AI now assists—or, in some cases, even replaces—human decision-making in matters of health. This represents an unprecedented development in human history.

In the era of AI-assisted health care, decision-making patterns can be categorized into 4 types based on the extent of AI involvement in the health care process (Figure 1): traditional physician-patient interaction (in which physicians make diagnostic and treatment decisions without AI support relying solely on patient-provided information and conventional examinations), patient-AI interaction (where patients use AI tools to support their own health care self-management decisions), physician-AI interaction (in which physicians use AI for clinical decision support), and physician-AI-patient collaborative interaction (where physicians and patients jointly use AI to facilitate shared decision-making throughout the diagnostic and treatment process) [7,8].

The transition from digital health to AI health represents a paradigm shift in health care, evolving from a physician-patient dyadic model to a triadic physician-patient-AI interaction model. This transformation constitutes an inevitable trend as it holds significant potential to address systemic challenges such as reducing unaffordable costs for patients, alleviating excessive workloads and physician burnout, and enhancing health care quality and operational efficiency [9]. By advancing toward a more integrated, intelligent, and human-centered system, this shift can ultimately contribute to better health outcomes and sustainable well-being [10].

However, the development of AI technology is uneven across countries, which has led to disparities in its application in the health care sector, resulting in inequalities in health services and health outcomes [11]. According to the Artificial Intelligence Index Report 2025, the United States leads in AI model development with 40 models in 2024 and a dominant cumulative output over the past decade. China produced 15 models in 2024, with its leading model (eg, DeepSeek) achieving near–state-of-the-art performance. Europe released only 3 models, continuing a declining trend, whereas regions such as the Middle East, Latin America, and Southeast Asia contributed sporadically and minimally [5]. In terms of research on AI applications in health care, the United States holds significant influence, and AI is widely used in specialized fields such as radiology and pathology, optimizing clinical decision support systems. Following the release of the DeepSeek model, many of China’s hospitals have deployed its public version [12]. In other countries, AI model development remains at an early stage and relies more heavily on international collaboration, resulting in only sporadic outputs; a similar pattern is seen in AI health applications, to which these countries have so far made only limited contributions [5].

In addition to gaps in AI research and health care deployment, access to and use of these technologies remain unequal across demographic groups. As the most advanced frontier of AI, large language models (LLMs) are predominantly controlled by a few leading companies and are close sourced to the public or health care providers. Given the restricted access to these LLMs, the resulting AI health applications are also typically released only to specific user groups by their developers. Similar to gaps observed in digital health [13-15], AI health inequalities stem from disparities in computational power (eg, graphics processing unit [GPU] chips), data resources (eg, standardized datasets), and algorithms (eg, LLM training frameworks). Gaps in access to AI health applications, differences in application skills, and uneven opportunities for use further exacerbate AI health inequity [16,17]. However, with DeepSeek leading a new wave of open-source initiatives [12], the widespread adoption of LLMs has the potential to fundamentally improve the foundation of AI health equity.

Why and how do open-source LLMs, compared to closed-source LLMs, improve AI health equity? Investigating this question helps policymakers, LLM companies and developers, hospitals, physicians, patients, and public health professionals understand the role and trends of AI in the health domain and actively promote the equitable access and application of open-source LLMs in health care. On the basis of a conceptualization of AI health and an analysis of the factors influencing AI health inequality, this study conducted a comparative analysis of the advantages of open-source LLMs over closed-source LLMs in improving AI health equity.. Specifically, using the health service triangle model as an analytical framework, this study compared open-source and closed-source LLMs across health service types, accessibility, quality, and cost. It argues that the local deployment of open-source LLMs offers unparalleled advantages in protecting patient privacy and data security, which in turn facilitates the development of more diverse health services, improves accessibility, enhances quality, and reduces costs. While open-source LLMs face challenges such as hallucinations, technology acceptance, and ethical issues, they are ultimately expected to improve AI health equity through technology sharing.

The development of AI has historically followed 2 distinct paths: closed-source and open-source. The fundamental tension between these approaches reflects a divergence between technological democratization and commercial monopolization. The closed-source camp advocates for maintaining proprietary control over technology to ensure service quality, achieve commercial returns, safeguard model security and intellectual property, and optimize model performance. In contrast, the open-source approach promotes the sharing of model algorithms, code, and data to engage global developers in improving the models, thereby accelerating innovation and broadening accessibility.

In the early stages of LLM development, many academic institutions and small to medium enterprises embraced open-source principles. However, as LLMs advanced, major technology giants and unicorns increasingly adopted closed-source strategies. Although some leading firms—such as Meta with its Llama series—have pursued open-source releases, their models often lag in performance compared to proprietary alternatives and have not significantly challenged the dominance of closed-source approaches. However, with the emergence of high-performance open-source models such as DeepSeek, the global AI industry is experiencing an unprecedented wave of open-source innovation. Table 1 provides detailed comparisons.

The openness of LLMs can be categorized into 3 levels: fully open-source, partially open-source, and planned open-source. Fully open-source models not only release model weights but may also include training code, datasets, inference deployment solutions, and more. They allow individuals, enterprises, and institutions to use the models freely, including for commercial purposes, without requiring additional authorization. DeepSeek’s series of models are fully open-source, permitting free use, local deployment, closed-source fine-tuning, and commercial applications. Fully open-source models have emerged as a key enabler for democratizing LLMs and fostering innovation in health applications. By lowering the barriers to accessing advanced LLMs, open-source initiatives allow diverse stakeholders—from regional hospitals to community clinics and individuals—to develop and deploy tailored health solutions. This accessibility helps reduce disparities in health care capabilities across different countries and demographic groups, thereby promoting health equity through broader and more equitable adoption of AI health technologies.

The direct factors influencing health include demographic and biological factors, environmental factors, behavioral factors, psychological factors, and health care factors [18,19]. The application of AI can impact most of these factors, thereby exerting a broad influence on health—a concept we define as “AI health,” denoting the application of AI technologies within the health sector to shape health outcomes. AI health equity involves multiple levels and dimensions, ranging from individual to regional scales and from behavioral to health care factors. Inequities in foundational resources (computational power, data, and algorithms) and throughout the AI life cycle (research and development, application, access, and use) within health care drive disparities in health outcomes. This constitutes AI health inequity. Thus, compared to closed-source LLMs, open-source LLMs demonstrate greater potential to reduce these disparities and contribute to enhanced AI health equity. At present, AI primarily affects health care services, improving health through health education and promotion, prevention, treatment, and rehabilitation. Given that the impact of AI on health is a relatively recent development and the significant influence of open-source LLMs on health only emerged in early 2025, this paper focuses specifically on the impact of LLMs on health service equity.

For the health service industry, ensuring data security and patient privacy remains one of the most important concerns when applying LLMs [20-22]. A significant advantage of fully open-source LLMs is that they can be deployed locally, enabling health care institutions to maximize the protection of patient privacy and data security [21,23-26]. Therefore, for both the demand side (eg, patients and health care providers) and supply side (eg, developers and institutions), open-source LLMs can help address health service inequalities often associated with closed-source LLMs. By overcoming these key barriers, open-source LLMs pave the way for enhancing the scope, quality, and affordability of health services, contributing to a more comprehensive improvement in AI health equity on a global scale.

In the field of health services, 2 models are widely recognized: one is the World Health Organization’s health system model focusing on the health care supply side [27], and the other is the model by Andersen [28] emphasizing the health care demand side. By integrating both supply and demand perspectives, Wang et al [29] developed a health service triangle model based on the essential common elements of these 2 frameworks to analyze online health services (for details, refer to Figure 2). Furthermore, this model can also be applied to evaluate the impact of open-source LLMs on AI health equity.

The health service triangle model highlights health services and their 3 indispensable fundamental elements: accessibility, quality, and cost. Health services form the core of the model, and the type of service determines its accessibility, quality, and cost. What kinds of health services can LLMs actually help provide? Without health services, there is no AI health—the type of health service is the foundation for evaluating AI health equity. For instance, emergency and urgent care services prioritize accessibility. If quality and cost are prioritized instead, delayed treatment could lead to severe health consequences. Accessibility includes geographic, temporal, economic, social, and eligibility dimensions [30,31]. Certain health services delivered via LLMs through online channels hold significant advantages in accessibility. Users can overcome spatiotemporal constraints, accessing services anonymously, inexpensively, or even free of charge around the clock [29].

Health service quality refers to the extent to which health care professionals and health care organizations deliver care to individuals and populations in ways that align with the best expected health outcomes achievable given current medical knowledge and technology [32]. LLMs have the potential to significantly enhance diagnosis, prescription, and treatment and are expected to address issues such as physician shortages, substandard care, and medical errors. The rapid rise in health care costs, which often makes care unaffordable for patients, has long been a fundamental barrier to the sustainable development of health systems. LLM technology, capable of efficiently processing medical images, assisting in diagnosis, discovering drugs, and improving medical education and research, is also seen as a promising tool to curb escalating medical expenses [9]. In terms of indirect costs—such as travel expenses—LLM-powered telemedicine can significantly reduce these burdens.

On the basis of the nature of health services and their triangular attributes, we established a comparative analysis framework for closed-source and open-source LLMs (for details, refer to Table 2). Compared to closed-source LLMs, open-source LLMs generate more positive social outcomes in terms of health service types, accessibility, quality, and cost—leading to a fairer distribution across regions and populations. In summary, the advantages of open-source LLMs over closed-source ones—in terms of health services and their 3 core attributes—enable AI to be widely used to meet various health service needs of both providers and users. This is particularly beneficial for remote and rural areas, financially constrained institutions, and privacy-sensitive applications. While closed-source LLMs can only offer limited generalized services targeted at users with payment capacity, open-source LLMs advance AI health equity by broadening access through more diverse and affordable services.

This study used a multimethod approach to compare open-source and closed-source LLMs across 4 dimensions of health services: types, accessibility, quality, and cost. Data were collected through web-based surveys; document research; and a systematic literature review, which included 31 AI model web pages (2020-2025), 7 DeepSeek technical documents, and 80 academic publications. The analysis quantified the relative strengths and weaknesses of each model type, demonstrating how the open-source paradigm promotes the use of AI-supported health services and contributes to improved health outcomes. The analysis shows that open-source LLMs are rapidly empowering both supply-side (eg, developers and institutions) and demand-side (eg, patients and health care professionals) actors, especially those in low- and middle-income countries (LMICs), to adopt AI in addressing their respective challenges. As AI-enabled health services become more widespread, this trend supports the achievement of AI health equity on a broader scale.

Since its emergence, AI technology has been applied in the health care domain, evolving from early expert systems to the current variety of diagnostic and treatment support systems, which primarily embody the physician-AI interaction pattern. However, the widespread and direct engagement of patients with AI technology became prominent only with the advent of LLMs. This trend accelerated significantly following the open sourcing of LLMs, facilitating the patient-AI interaction pattern, where patients began using these models extensively for consultation, self-diagnosis, and even prescription information gathering. The proliferation of these interaction patterns has propelled AI health into a widespread phenomenon, marking a new and distinct phase in the evolution of digital health. In terms of accessibility, quality, and cost, the greatest advantage of open-source LLMs over closed-source alternatives lies in their ability to make both the physician-AI interaction and patient-AI interaction patterns more widespread and equitable.

Open-source LLMs not only promote AI health equity in traditional dimensions—such as across nations, regions, and populations—but also, by enabling shared access for both health care providers and patients, foster equity in a new dimension: a more balanced clinician-patient relationship within the physician-AI-patient interaction pattern. In traditional health care systems, only medical professionals hold the authority to make diagnostic and treatment decisions, whereas patients are typically excluded from participating in choices concerning their own health and disease management. A significant information asymmetry exists between physicians and patients regarding health and medical knowledge, resulting in highly unequal power dynamics during clinical decision-making.

The emergence of LLMs, especially open-source LLMs, has disrupted this imbalance. A growing number of patients now upload their laboratory results, medical images, and even full diagnostic records to LLMs seeking diagnostic interpretations and management recommendations based on their symptoms, history, and clinical data. Although the output of these models is intended for reference only, their high diagnostic accuracy—often surpassing that of average physicians—has led some patients to use LLM-generated insights to facilitate discussions with their physicians regarding diagnosis and treatment plans. Others rely on such suggestions to evaluate the appropriateness of the care recommended by their clinicians. On the health care supply side, medical providers have already begun integrating LLMs into professional workflows, making decisions based on AI-generated assessments or even fully adopting AI-proposed solutions. Introducing open-source LLM–supported diagnostics into medical decision-making enables shared decision-making between clinicians and patients based on AI-informed insights. This transforms the conventional care paradigm, leading to a more equitable relationship between patients and providers throughout the diagnostic and therapeutic process.

The technological democratization driven by open-source LLMs holds significance not only for countries such as China that are leading the open-source movement but also for global health. A core objective of global health is health equity, and addressing the growing disparities in health outcomes worldwide is a shared challenge for all nations. Open-source LLMs substantially enhance the accessibility of AI technology for both supply- and demand-side actors in health care. As long as patients have basic internet-connected devices such as smartphones, they can access open-source LLMs. Similarly, health care providers such as hospitals can deploy suitable models if they possess fundamental computing resources. This is particularly transformative for hospitals, physicians, and patients in low- and middle-income countries as open-source LLMs can effectively improve the accessibility and quality of their health services. Countries previously marginalized by expensive commercial systems can now leverage mobile apps to facilitate initial self-diagnosis [50,51], thereby advancing progress toward universal health coverage.

While open-source LLMs hold significant promise for improving AI health equity, their application in the medical field still faces considerable challenges. Open-source LLMs are prone to a certain degree of AI hallucination, making it crucial to ensure the accuracy and reliability of the medical advice they generate. Some patients may engage in self-medication based on LLM-generated recommendations [50], and errors in such advice could be life-threatening. Although open-source LLMs emphasize that their suggestions are “for reference only” and advise users to “follow professional medical guidance,” legal and ethical questions regarding accountability in cases of medical errors or malpractice remain unresolved [37]. Regarding the openness of LLMs and their diverse users, corresponding ethical and governance strategies should be established. For instance, in the case of deploying fully open-source LLMs, the responsibility for model custodianship, updates, and accountability for application outcomes lies with the deployer rather than the developer.

The comprehensive and standardized recommendations generated by open-source LLMs can mitigate information asymmetry between physicians and patients, yet they may also inadvertently undermine trust in clinical authority. Moreover, integrating even the smallest-parameter open-source LLMs into existing health information systems demands dedicated staff training and infrastructure investment. Meanwhile, the deployment of variably scaled models—such as 70B vs 671B parameter versions—contingent on local hardware capacity may perpetuate or even exacerbate disparities in AI adoption. To address these challenges, global policymakers must collaborate with stakeholders across sectors to establish robust governance frameworks for medical AI. Such frameworks should proactively tackle technical, financial, and ethical barriers, ensuring that AI-driven health services are effectively deployed worldwide and equitably accessible, thereby advancing health outcomes for all.