Lunit develops LearningUnit 2.0 medical AI foundation model

Lunit has completed the development of LearningUnit 2.0, a lightweight medical-specific foundation model, following a 10-month project led for the Ministry of Science and ICT. The model features 30 billion parameters, with 5 billion active, and has been released as open source on Hugging Face.

Designed with an agent structure, the model moves beyond simple chatbots to support complex medical workflows, including clinical decision support, emergency room triage, multidisciplinary care, adverse drug reaction assessment, and insurance claims.

During a six-month validation period involving nine medical institutions, including Korea University Medical Center and Ilsan Hospital, the model demonstrated high clinical accuracy. In multidisciplinary care, 86.2% of the AI-suggested treatment plans were deemed clinically valid. The model achieved a 98% agreement rate with specialists in emergency room triage and adverse drug reaction causality assessments. Additionally, 85.6% of its responses in insurance review and claims were rated as practical for professional use.