SNUH unveils Korea’s first homegrown medical LLM with 86.2% exam accuracy

SOUTH KOREA – Seoul National University Hospital (SNUH) has achieved a groundbreaking milestone by developing what could be the first medical large language model (LLM) entirely based in South Korea.

This development marks a significant advancement in the application of artificial intelligence in healthcare and represents a strategic step in tailoring medical AI to meet local linguistic, legal, and clinical standards.

The ambitious project began in March 2024, spearheaded by SNUH’s research team. The team collected and compiled over 38 million clinical texts, including a wide range of hospital documentation such as inpatient and outpatient records, surgical and prescription notes, and nursing reports.

All data underwent strict anonymization and de-identification protocols to create a safe and robust training dataset for model development.

Building on this foundation, SNUH launched the second phase of development in early 2025 by introducing department-specific knowledge bases.

These knowledge repositories incorporated Korean medical laws, journal abstracts, national treatment guidelines, standardized medical terminologies, and a comprehensive abbreviation dictionary.

The team then utilized retrieval-augmented generation (RAG) to fuse these datasets into a single, highly specialized model capable of nuanced understanding and response generation in Korean medical contexts.

When tested using questions from the Korean Medical Licensing Examination spanning the past three years, the model achieved an impressive accuracy rate of 86.2%, surpassing the national average score of 79.7%.

Beyond clinical exams, the model demonstrated powerful translation capabilities, reportedly handling texts of up to 50,000 words at once, indicating its potential utility in real-world medical documentation and research.

According to SNUH, most existing medical LLMs such as Google’s Med-PaLM 2 and Microsoft’s LLaVA-Med are not adequately equipped to handle Korean medical language and standards.

These models, developed for Western clinical contexts, often misinterpret Korean-specific terminology, guidelines, and legal frameworks.

In response to these limitations, SNUH aimed to build a tailored solution that supports Korean healthcare professionals more effectively.

Over the coming year, the research team plans to focus on validating the model’s safety, accuracy, and clinical utility through rigorous internal testing.

Only after this validation phase will the model be applied in clinical settings to assist with tasks such as diagnosis support, medical documentation, and research summarization.

Future plans also include broadening the model’s capacity to process complex datasets and expand its reach across various departments and specializations.

This project is part of a larger AI-driven movement within South Korea’s medical landscape. In late 2024, PhynX Lab, backed by the SK Group, introduced Cheiron, an LLM-based healthcare search platform aimed at pharmacists and pharma researchers.

Meanwhile, Korea University Anam Hospital is developing its own LLM, expected to be piloted soon, and Asan Medical Center has already deployed an LLM-powered voice scribe across 16 departments. SNUH is concurrently working on additional AI tools designed to ease administrative burdens on clinicians.

These include a multimodal AI system to automatically generate patient summaries, an insurance AI for streamlining claims processing, and a literature curation AI that recommends up-to-date academic papers to researchers.

“Through the development of this Korean-style medical LLM, we have opened a new chapter in medical innovation by maximising the work efficiency of medical staff and providing faster and more accurate medical services to patients,” said Dr. Kim Young-tae, President and CEO of SNUH.