DeepSeek Files IPO Prospectus Amid Surge in Healthcare AI Adoption

No technology has been able to penetrate various industries as rapidly as large language models after breaking out of their initial niche. Yet, before they could find viable commercial pathways, the industry had already become mired in intense competition over parameters, costs, and performance, leading to a deep entrenchment in computational power “stacking.”

This January, the emergence of DeepSeek-R1 rewrote the rules of the game that had been dominated by GPT models over the past year. With its innovative model architecture and training optimization strategies, DeepSeek demonstrated to the industry that high-performance general-purpose models can be achieved even with a limited scale of parameters.

Beyond breaking the “computing power monopoly,” DeepSeek’s innovative designs, such as Parameter-Efficient Fine-Tuning (PEFT) and Mixture of Experts (MoE) architecture, have also successfully lowered the “entry barrier” for large models.

The combination of low costs and the “Made in China” label has prompted a rapid deployment by numerous top-tier hospitals and cutting-edge medical technology companies across China, with even the National Healthcare Security Administration prominently announcing its integration of DeepSeek, once again thrusting large language models into the spotlight.

Is It Mere Bandwagoning or a New Path? VCBeat recently engaged in dialogues with tech-medical companies that have integrated DeepSeek, providing point-by-point answers to three questions: “The Real Value of DeepSeek in the Healthcare Sector,” “Modes of DeepSeek Application in Hospitals,” and “The Current Status of Medical Scenario Application Development Based on DeepSeek-R1.”

Long before the advent of DeepSeek-R1, hospitals in China had already deployed general-purpose models, proactively embarking on their exploration of generative AI.

Since clinically relevant data cannot leave the hospital premises, large language models at the time could only be deployed through on-premises integration. The challenge lies in the fact that most hospitals’ IT infrastructure is primarily based on CPUs designed for general-purpose computing, with few hospitals possessing GPU resources tailored for graphics processing and parallel computing, making it difficult to provide sufficient computational power.

The Dilemma of Computing Power Is Tied to Cost. Among numerous hospitals, the leading institutions have the financial capacity to invest heavily in comprehensive GPU infrastructure, fully deploying general-purpose models within their premises to serve hospital-wide systems; a small minority can streamline these models to benefit specific departments.

When the vast majority of medical institutions are unable to freely deploy large language models (LLMs) or develop related clinical applications, companies operating in the medical LLM sector also struggle. Facing a lack of sufficient buyers, they find it difficult to sustain high levels of R&D investment in the direction of large language models.

The emergence of DeepSeek-R1 has disrupted this status quo. Leveraging an innovative architecture and open-source code, it fundamentally resolves the cost issues associated with the deployment and operation of general-purpose models.

Wu Di, CEO of Fuxin Kechuang, stated: “Since DeepSeek-R1 adopts a Mixture of Experts (MoE) architecture, it activates only approximately 37 billion parameters (out of a total of 671 billion) during each inference. This avoids the high computational costs associated with traditional dense models, which require full parameter activation. Theoretically, this approach maintains inference accuracy while reducing computational power consumption by more than 40%. Furthermore, if enterprises need to scale up the model, they can enhance its capabilities without having to increase computational investment linearly.”

 Comparison of Capabilities: DeepSeek, GPT o1, and GPT o3 mini

(The input price only reflects the standard-hour pricing under Cache Hit conditions; data sources: VCBeat, DeepTrans AI)

More importantly, DeepSeek is released under the highly permissive MIT License, which allows users to deploy it locally and freely use, copy, modify, and distribute the software. This also encourages enterprises to adopt and integrate it into their products, fostering collaboration and innovation, thereby driving the development of the entire ecosystem.

This open ecosystem enables ordinary medical institutions to develop large medical models that are better aligned with real-world application scenarios, based on their specific business needs. By deploying smaller models with fewer than 100 billion parameters obtained through distillation, many primary healthcare facilities can run these models smoothly using their existing integrated graphics cards.

"In our communications with regional medical institutions, we found that their needs are even more clearly defined: they hope to leverage DeepSeek’s reasoning capabilities at the primary care level, where there is the greatest shortage of physicians capable of handling complex cases."

Overall, the value of DeepSeek-R1 lies in lowering the barrier to entry for large language model (LLM) applications, opening up new markets for practical implementation, and accelerating the emergence of vertical-specific applications. In this process, this emerging model has paved the way for the commercialization of medical LLMs.

As the number of hospitals planning to deploy large language models (LLMs) and physicians individually engaged in LLM development continues to grow, numerous upstream enterprises in the medical IT industry have become increasingly active.

According to Zhao Daping, CTO of Winning Health, following the emergence of DeepSeek-R1, mainstream deployment models in China can be broadly categorized into three types. First, hospitals can rapidly download the model from cloud or source endpoints and complete deployment quickly; this approach is primarily suitable for large hospitals that already possess GPU hardware. If a hospital lacks the necessary GPUs for computation, it can rent cloud-based resources. Additionally, some private hospitals opt for a subscription-based deployment model, mainly serving specific departments.

Moreover, the current trend has spurred the emergence of numerous companies manufacturing all-in-one large language model appliances. However, in Zhao Daping’s view, for hospitals to achieve effective operation of large models, they must first integrate these models with their existing hospital information systems; secondly, the information systems themselves should adopt intelligent architectures that support AI operations as much as possible.

After all, although large-model appliances can achieve a certain level of interactive capabilities through external integrations, they struggle to facilitate comprehensive data exchange with the dozens of legacy systems already deployed in hospitals. Unless an integrated “model + application” solution is implemented, it will be difficult to meet the diverse needs of healthcare institutions.

So, how should hospitals deploy large language models (LLMs) under ideal conditions? Zhao Daping believes that as LLMs become increasingly integrated, future hospital configurations will inevitably adopt a diversified hybrid approach. “Hospitals may deploy one general-purpose LLM alongside several smaller, specialized models tailored to specific clinical departments. The LLM would handle complex interactive scenarios requiring reasoning, critical thinking, and diagnostic support, while the smaller models would be applied in scenarios emphasizing rule-based operations, decision-making, correction, and simple content generation. This strategy ensures both cost-effectiveness and operational efficiency while meeting diverse clinical needs.”

"Extending this further, there are numerous mobile scenarios within hospitals. If we can deploy small models on smartphones, a significant portion of tasks in current medical workflows could be shifted to mobile devices, greatly enhancing healthcare efficiency."

Revisiting Physicians and Other Individuals Seeking to Proactively Develop Clinical Applications

As DeepSeek surged in popularity, a flood of tutorials emerged, encouraging users to independently configure and train models. However, in the medical field,Although the emergence of DeepSeek has lowered various barriers to model training, localizing the training of private models still requires researchers to possess a certain level of technical expertise, as it involves five key steps: data preparation and processing, model selection and configuration, model training, model evaluation and tuning, and model deployment and integration.

“Many current large language model (LLM) applications are underdeveloped. After purchasing GPUs and configuring models, many hospital research institutes aim to immediately build applications for specific scenarios, but they often find themselves lacking the necessary development capabilities in practice. To achieve widespread adoption among individual physicians and thereby generate research outcomes, we must wait for service providers to upgrade their user interfaces (UIs) and further simplify the development pathway for LLM-based applications.”

In other words, the joint development of vertical models by enterprises and medical institutions remains the central theme of medical AI.

Although DeepSeek-R1 has achieved large-scale deployment in the healthcare sector, its relatively short time since launch means it has not yet expanded beyond the existing application scope of large language models. Instead, it focuses more on reducing deployment and training costs and improving text processing efficiency. In the initial phase, a batch of large model enterprises specializing in internet healthcare were the first to benefit.

For instance, Tencent Health leveraged Tencent Cloud to integrate the DeepSeek series and, in combination with its self-developed Hunyuan large language model, rapidly iterated on medical services such as intelligent triage, pre-consultation, health Q&A, imaging report interpretation, and quality control. This effort accelerated the upgrade of intelligent applications for more than 1,000 hospitals across China.

Currently, Tencent’s “Shenzhen Medical Insurance” app has integrated the latest large AI models into its intelligent customer service system. Users can freely choose between DeepSeek, which excels in reasoning, or Tencent Hunyuan, which offers multi-dimensional understanding of queries. Whether consulting on complex policies such as “how maternity allowances are calculated” or asking professional questions like “how specific outpatient diseases are certified,” the integrated large model can provide precise and thoughtful answers tailored to the user’s specific insurance status, thereby helping users better understand their inquiries while responding.

As DeepSeek’s accumulated medical data continues to grow, its application advantages in hospital settings are gradually becoming apparent. Benefiting from significantly reduced requirements for prompt engineering and empowered by chain-of-thought technology, DeepSeek effectively enhances the transparency and interpretability of AI in clinical diagnosis, while enabling doctors to communicate more efficiently with the model.

For example, when doctors previously used large language models to generate surgical plans, they needed to provide complete and clear details such as past medical history and surgical conditions. In contrast, with DeepSeek, only key information needs to be input, as the model autonomously fills in the relevant details during its “Think” process.

Furthermore, medical reasoning emphasizes an evidence-based process. DeepSeek not only provides effective diagnostic and treatment recommendations but also elaborates in detail on the underlying reasoning, including diagnostic criteria, medication choices, and recommended tests. This transparency significantly alleviates physicians’ skepticism toward AI systems, offers a clear basis for doctor-patient communication, and thereby promotes broader clinical adoption of AI technologies.

“Many doctors pay close attention to the model’s ‘thinking’ process; they briefly review DeepSeek’s logic. This is a crucial interaction that helps build trust among physicians.”

To date, numerous hospitals have deployed large language model (LLM)-based applications. Taking medical documentation as an example,Fuxin Kechuang, Winning Health, and other companies have all developed similar applications. TakingTaking Fuxin Kechuang as an example, the company has deployed AI-powered generative electronic medical record (EMR) systems across multiple outpatient and inpatient scenarios at hospitals such as Wuhan Union Hospital and Zhongnan Hospital of Wuhan University, aiming to improve physicians’ efficiency in medical documentation.

In traditional outpatient consultations, a single patient visit is typically allocated 10 minutes, with approximately 5 minutes spent on documenting electronic medical records (EMRs), 3 minutes on prescribing medications and ordering tests, and an average of only 2 minutes dedicated to actual clinical inquiry. With the integration of AI, real-time doctor-patient conversations are captured and converted into standardized medical terminology, automatically generating EMRs according to outpatient templates, thereby eliminating the time previously required for manual documentation.

“Based on a physician seeing 50 patients per day, at least one hour of time spent on medical record documentation can be saved daily. If hospitals allocate this saved time to consulting more patients, large language models can generate tangible economic value for healthcare institutions.” Therefore, in Wu Di’s view, this is currently the highest-value scenario that is relatively easy to implement.

Since the DeepSeek model itself has not been trained on CT or MR imaging data, enterprises developing related applications need to establish their own imaging datasets and build models accordingly. Consequently, research based on the DeepSeek large language model in the field of medical imaging is relatively scarce compared to various text-based tools.

Currently, DeepWise has conducted preliminary explorations of DeepSeek at the internal tool level. For instance, they have applied DeepSeek to multimodal standardization and enhancement of imaging data. By leveraging image data alongside non-image metadata (such as EMR, HIS/RIS, and DICOM headers, which contain abundant textual information), they improve the consistency of imaging content and nomenclature. This optimization enhances downstream applications—for example, more accurate and consistent hanging protocols can improve physicians’ efficiency.

In terms of quality control data analysis, DeepWise Medical is leveraging large language models to enhance medical imaging quality control, anomaly detection capabilities, and interactive workflow issue resolution.

It is worth noting that while imaging studies based on DeepSeek remain limited, the industry has achieved substantial research outcomes in large imaging models. Some enterprises have established foundational imaging models based on architectures such as GPT, and clinical trials have confirmed that large language models (LLMs) improve both the accuracy and efficiency of medical image diagnosis. As DeepSeek’s capabilities continue to strengthen, these companies may gradually transition to domestically developed general-purpose models.

Beyond Hospital Settings: Drug R&D as a Key Arena for Large Language Models

Currently, DeepWise Medical is piloting the use of DeepSeek to address medical imaging standardization, thereby better resolving issues such as imaging data quality control in pharmaceutical R&D trials. According to Gong Enhao, CEO of DeepWise Medical, the company has signed agreements with a number of international pharmaceutical companies to optimize their existing imaging trial data from ongoing R&D projects.

Some other models, although not using DeepSeek, have also adopted similar innovative technologies.

For example, BioMap’s xTrimo series of large models also adopts the Mixture of Experts (MoE) framework. Its V3 version can process seven major modalities of data, including DNA, RNA, proteins, cells, compound–protein interactions, protein–protein interactions, and life systems. This enables full-scale modeling from base pairs to cell clusters, thereby empowering scientific research in areas such as antibody and cell and gene therapy drug development, target discovery, and microbiology.

However, it should also be noted thatWhether it involves empowering healthcare institutions or pioneering drug R&D, developers using large language models like DeepSeek are mostly upgrading existing scenarios rather than creating applications that disrupt them—let alone driving true innovation. Fortunately, DeepSeek-R1 has been available for less than two months; as time goes on, we may well witness surprising breakthroughs from medical AI.

Although the emergence of DeepSeek-R1 has significantly advanced the depth of large language model (LLM) applications in the medical field, rationally speaking, it will still take considerable time before LLMs can be routinely deployed in daily hospital operations.

First, solving complex problems requires large models to integrate patients’ multimodal data and perform comprehensive inference, much like physicians do. However, during the “Think” phase, DeepSeek often falls into potentially infinite loops, generating numerous responses irrelevant to the question at hand. In serious, high-frequency domains such as healthcare, these instances of hallucination must be eliminated to enable scalable deployment.

Second, DeepSeek’s “domestically produced” identity makes it more favored by domestic medical institutions; however, to achieve large-scale application, it must still comply with medical data privacy and security regulations. Therefore, DeepSeek needs to implement more robust data de-identification and encryption technologies to ensure the safety of patient data.

Third, DeepSeek addresses the product quality and performance shortcomings of previous large language models but has yet to identify a “killer application” that would drive healthcare institutions to proactively pay for its services. Currently, the willingness to pay for AI remains tied to user perception and whether the product itself can genuinely reduce costs, improve efficiency, generate revenue, and provide empowerment. Therefore, for DeepSeek to achieve large-scale implementation, it must first enhance acceptance among hospitals and physicians, and second, further advance beyond traditional AI capabilities. Regarding the question of who will pay, a review of AI’s development over the past decade suggests that primary care settings have a greater need for large language model support than tiered hospitals.

Fourth, DeepSeek’s technological breakthroughs are not irreplicable. Today, certain versions of GPT have significantly reduced model training costs, approaching DeepSeek’s level, while continuously enhancing their logical reasoning capabilities. This necessitates that DeepSeek further consolidate its advantages and deliver tangible results in addressing real-world clinical challenges.

Despite the numerous challenges, we can still identify many positive developments. After all, the participation of a large number of medical enterprises and healthcare institutions will inevitably give rise to more vertical applications, thereby expanding the possibilities for the commercialization of large language models.

Meanwhile, the inherent potential of models such as DeepSeek should not be overlooked. Given the current pace of iteration for large language models (LLMs), general-purpose models undergo a comprehensive update cycle every three months. It is possible that sometime in 2025, we will witness a particular LLM emerge as a leader, systematically addressing the aforementioned challenges and, together with numerous health-tech enterprises, ushering in a new era for medical large language models.

Acknowledgments

Wu Di, CEO of Fuxin Sci-Tech Innovation

Zhao Daping, CTO of Winning Health

Chen Xu, Head of the Artificial Intelligence Laboratory at Winning Health

Gong Enhao, CEO of DeepWise Medicine

Wu Zhigang, General Manager of Tencent Health User Platform
Wu Xian, Head of the Tianyan Research Center at Tencent Youtu Lab
Support for This Article

VCBeat’s VBEF Top 100 Future Healthcare & Pharma Expo has accompanied us for a decade, during which we have jointly witnessed the Chinese healthcare industry’s stride toward “newness.” The sector continues to surge with robust innovative power and showcases its unique charm on the global stage. In May 2025, we will continue to uphold this mission by hosting the 2025 VBEF Top 100 Future Healthcare & Pharma Expo in Suzhou. During the conference, we will hold the Medical AI Large Model Application Innovation Forum at 2:00 PM on May 9. Readers interested in AI in healthcare are welcome to scan the QR code to register for the event.