SB Intuitions, Juntendo University, SoftBank, and HEMILLIONS have launched a joint research project utilizing “Med-Sarashina,” a Japanese domestic Large Language Model (LLM) specialized for the medical field. This initiative addresses the urgent need in Japan for secure AI solutions that comply with local medical regulations and the unique linguistic nuances of the Japanese healthcare system.

Launch of Joint Research and Organizational Roles in Japan

A Collaborative Project Commenced in June 2026

On August 10, 2026, SB Intuitions, a subsidiary of SoftBank Corp., announced that it began a joint research project in June 2026 with Juntendo University, SoftBank, and HEMILLIONS (a medical information system developer). This project focuses on the social implementation of generative AI in Japan’s clinical, research, and educational fields using the medical-specific LLM “Med-Sarashina“. While digital transformation is a priority in the Japanese medical sector, the need for high expertise and strict information management has made implementation challenging. This partnership represents a landmark effort to solve these technical and ethical hurdles through industry-academia collaboration within Japan, ensuring that AI can be safely integrated into the busy workflows of Japanese hospitals.

Strategic Partnership Among Four Organizations

In this joint research, each of the four Japanese organizations plays a distinct and crucial role to ensure the project’s success. Juntendo University oversees the entire research project, utilizing its extensive medical expertise to build high-quality clinical reasoning datasets and perform clinical validation and evaluation of the AI models. SB Intuitions provides the “Med-Sarashina” model and offers technical advice on dataset construction based on its experience in developing Japanese-specific LLMs. SoftBank supports the practical implementation and evaluation of the AI model while exploring sustainable business models for social implementation in Japan. Finally, HEMILLIONS is responsible for building and operating the secure medical data platform, ensuring it complies with the strict “3 Ministries 2 Guidelines” set by the Japanese government. This integrated structure allows for a seamless transition from basic research to real-world validation in Japanese medical settings.

Technical Foundation of the Medical-Specific LLM “Med-Sarashina”

Full-Scratch Development Independent of Foreign Foundational Models

The primary technical strength of “Med-Sarashina” is its full-scratch development, meaning it was trained from the ground up rather than being a modified version of an existing foreign foundational model. Many generative AI services currently used in Japan rely on overseas vendors, raising concerns about data being processed on foreign servers or used for unintended training purposes. By developing a model entirely within Japan, SB Intuitions ensures data transparency and maintains control over the entire training process. This approach directly addresses the concerns of Japanese medical institutions regarding the handling of sensitive patient information. Furthermore, this independent development allows the model to accurately capture Japanese business customs, subtle linguistic nuances, and the specific terminology used in the Japanese medical context, which are often mistranslated or misunderstood by foreign-made AI.

Technical Architecture for Enhanced Medical Reasoning

To refine the model for clinical use, the project involves fine-tuning “Med-Sarashina” with clinical reasoning datasets supervised by Juntendo University. This process aims to develop specialized models like “Med-Sarashina-Juntendo,” which possess a deep understanding of complex Japanese medical contexts. The research also explores integrating advanced technologies such as Retrieval-Augmented Generation (RAG) and Knowledge Graphs to move beyond simple text generation. These technologies enable the AI to provide evidence-based diagnostic support and assist in the creation of highly technical medical documents. By structuring complex medical data, the team aims to create a tool that does not just answer questions but actively supports the clinical thinking process of Japanese physicians.

Addressing Challenges Unique to the Japanese Healthcare Environment

Overcoming the Limits of Localization in Foreign AI Models

General-purpose foreign LLMs are primarily trained on English-language data, which can lead to inaccuracies regarding Japanese-approved drugs, local medical customs, and specific legal regulations. For instance, an AI might recommend a medication that is unapproved in Japan or suggest a treatment plan inconsistent with the Japanese public insurance system, a phenomenon known as “hallucination”. A Japanese medical-specific model can directly incorporate Japanese clinical guidelines and the latest local findings, providing far more reliable support. Consequently, a domestic LLM trained on data deeply rooted in the Japanese medical environment is seen as essential for reducing risk and providing appropriate assistance for the daily decisions made by Japanese healthcare professionals.

Compliance with Strict Japanese Data Privacy Regulations

In Japan, medical information is classified as “Special Care-Required Personal Information,” requiring a higher level of protection than standard personal data to prevent discrimination or prejudice. While the Ministry of Health, Labour and Welfare (MHLW) provided official guidance (Q&A 企Q-26) in May 2025 regarding the use of generative AI, the location of data storage remains a significant concern for Japanese hospitals. This project operates on the premise of using Japanese data centers and adhering to the “3 Ministries 2 Guidelines,” a rigorous security standard for medical information systems. Specifically, the research validates advanced operational methods such as Zero Data Retention (ZDR), where input data is immediately discarded and not used for further training or saved in logs, meeting the legal and technical requirements of the Japanese medical sector.

Practical Scenarios in Clinical and Educational Settings

Standardizing Japanese Medical Education Through AI Infrastructure

A major pillar of this joint research is improving the clinical reasoning abilities of Japanese medical residents. The AI can present clinical cases, evaluate a resident’s diagnostic process or inquiry strategy, and provide immediate feedback, thereby standardizing and improving the quality of medical education in Japan. This also serves to reduce the heavy teaching burden on senior Japanese physicians. In the future, the data accumulated through this system could help visualize the growth rate of young Japanese doctors and lead to more consistent educational outcomes across different Japanese medical institutions. The following diagram illustrates the expected impact of AI agents in the medical field.

Please refer to the figure below.

Reducing Physician Overtime via Automated Documentation

Japanese medical professionals face significant challenges with long working hours, often driven by the massive volume of administrative work, such as writing referral letters and discharge summaries. According to MHLW surveys, “document preparation” is a leading cause of overtime in Japanese hospitals. This joint research focuses on developing tools where the AI summarizes information from electronic health records to automatically generate drafts of these documents. By automating these clerical tasks, the project aims to support the Work-Style Reform for Physicians in Japan, allowing doctors to dedicate more time to direct patient care. The team is gathering direct feedback from Japanese medical staff to design a user experience that integrates smoothly into actual workflows, ensuring the tools remain a sustainable part of hospital operations.

Future Prospects and the Evolution of Japanese Medical AI

Accelerating Implementation Across the Japanese Healthcare Sector

Moving forward, SB Intuitions plans to further refine the safety and accuracy of its models through continued demonstrations at Juntendo University before proceeding with a phased social implementation. The insights gained from this project are expected to be shared with other Japanese medical institutions and regional medical networks, acting as a catalyst for Medical DX (Digital Transformation) across Japan. Successfully establishing these case studies at a premier institution like Juntendo University sends a powerful message to the entire Japanese healthcare industry, lowering the psychological and technical barriers for other hospitals to adopt AI.

Synergies Within the SoftBank Group’s Holistic AI Strategy

The broader SoftBank Group is pursuing a multi-faceted AI medical strategy, including joint research with Chugai Pharmaceutical on AI for clinical development and a partnership with the U.S.-based Tempus AI to utilize real-world multi-modal data. The research into medical-specific LLMs by SB Intuitions will likely serve as a core component of this ecosystem, potentially contributing to a next-generation “Medical AI OS” that supports everything from drug discovery to personalized diagnosis and treatment in Japan. If SB Intuitions’ top-tier Japanese language performance is fused with the drug discovery DX of Chugai Pharmaceutical and the data analysis technology of Tempus AI, Japan could evolve into a unique and powerful player in the global medical AI market.

Reference Pages

• 【国産医療特化型LLM「Med-Sarashina」を活用した共同研究を開始】https://www.sbintuitions.co.jp/news/press/20260810_01/

• 【中外製薬、ソフトバンク、SB Intuitionsの3社が生成AIの活用で臨床開発業務を革新し、新薬開発のスピードアップを目指す共同研究に向けた基本合意を締結】https://www.chugai-pharm.co.jp/news/detail/20250130153000_1461.html

• 【医療現場の事務作業を支援する高性能な日本語LLMを開発しました】https://www.t.u-tokyo.ac.jp/press/pr2026-05-28-001

【#MedicalAI #GenerativeAI #DomesticLLM #JuntendoUniversity #SoftBank #MedicalDX #Med-Sarashina】