Lunit Completes LearningUnit 2.0

Lunit said on November 6 that it had completed the Ministry of Science and ICT’s “AI Specialized Foundation Model” project and finished developing LearningUnit 2.0, a model tailored to healthcare. The lightweight model has 30 billion parameters, of which 5 billion are activated, and Lunit released the results as open source on Hugging Face. After being selected in October last year to lead a national strategic project, Lunit led a consortium of 23 industry, academic, research and medical institutions for about 10 months, developing the model and testing it in clinical settings.

LearningUnit 2.0 was designed to understand medical workflows and provide support at each step, beyond answering healthcare questions like a chatbot. Its agent structure (a system that performs tasks through coordinated steps) analyzes data, presents its reasoning and connects to follow-up tasks. The development focused on clinical decision support, emergency patient triage and initial assessment, multidisciplinary care, causality assessment of adverse drug reactions, and insurance review and claims.

For about 6 months, Lunit tested the model using real-world data at 9 medical institutions nationwide, including National Health Insurance Service Ilsan Hospital, Konyang University Hospital and Korea University Medical Center. In complex multidisciplinary care, clinicians rated 86.2% of AI-proposed treatment plans as clinically valid. For emergency severity classification (KTAS) and adverse drug reaction causality assessments (ADR), results were within one level of specialist judgments in 98% of cases for each. Assessors also rated 85.6% of the model’s insurance review and claims responses as usable in practice.

The model met target performance across all official benchmarks covering medical knowledge reasoning, factual accuracy of citations and stability in medical conversations; a third party verified the results. Across 17 benchmarks, including areas outside the project’s performance targets, it recorded average accuracy comparable to frontier models from global technology companies. Lunit said the 30-billion-parameter LearningUnit 2.0 was more accurate than GPT-OSS, an open-source model with 120 billion parameters, and performed at a similar level to GPT 5.2, estimated to be hundreds of times larger. Lunit CEO Se Bumseok said the model is infrastructure intended to support hospitals’ everyday work across departments, rather than a single disease or function, and that the company would develop it into technology that helps medical staff and patients, following its performance checks in domestic healthcare settings.

The reference for this summary is a public press release dated November 6, 2026. That is the article’s publication date; the article says only that the project completion was announced on the 6th. The article does not say that AI-proposed treatment plans or classification results replace clinicians’ judgments. It reports results from real-world trials and benchmark evaluations, but gives no details on the size of the trial datasets or the evaluation methods. Lunit released the model, and CEO Se Bumseok said the company would develop it as infrastructure for hospital-wide support. The reported hospital trial figures were 86.2% for multidisciplinary care, 98% agreement each for KTAS and ADR, and 85.6% for insurance review and claims.

Lunit said on November 6 that it had completed the Ministry of Science and ICT’s “AI Specialized Foundation Model” project and finished developing LearningUnit 2.0, a model tailored to healthcare. The lightweight model has 30 billion parameters, of which 5 billion are activated, and Lunit released the results as open source on Hugging Face. After being selected in October last year to lead a national strategic project, Lunit led a consortium of 23 industry, academic, research and medical institutions for about 10 months, developing the model and testing it in clinical settings.

LearningUnit 2.0 was designed to understand medical workflows and provide support at each step, beyond answering healthcare questions like a chatbot. Its agent structure (a system that performs tasks through coordinated steps) analyzes data, presents its reasoning and connects to follow-up tasks. The development focused on clinical decision support, emergency patient triage and initial assessment, multidisciplinary care, causality assessment of adverse drug reactions, and insurance review and claims.

For about 6 months, Lunit tested the model using real-world data at 9 medical institutions nationwide, including National Health Insurance Service Ilsan Hospital, Konyang University Hospital and Korea University Medical Center. In complex multidisciplinary care, clinicians rated 86.2% of AI-proposed treatment plans as clinically valid. For emergency severity classification (KTAS) and adverse drug reaction causality assessments (ADR), results were within one level of specialist judgments in 98% of cases for each. Assessors also rated 85.6% of the model’s insurance review and claims responses as usable in practice.

The model met target performance across all official benchmarks covering medical knowledge reasoning, factual accuracy of citations and stability in medical conversations; a third party verified the results. Across 17 benchmarks, including areas outside the project’s performance targets, it recorded average accuracy comparable to frontier models from global technology companies. Lunit said the 30-billion-parameter LearningUnit 2.0 was more accurate than GPT-OSS, an open-source model with 120 billion parameters, and performed at a similar level to GPT 5.2, estimated to be hundreds of times larger. Lunit CEO Se Bumseok said the model is infrastructure intended to support hospitals’ everyday work across departments, rather than a single disease or function, and that the company would develop it into technology that helps medical staff and patients, following its performance checks in domestic healthcare settings.

The reference for this summary is a public press release dated November 6, 2026. That is the article’s publication date; the article says only that the project completion was announced on the 6th. The article does not say that AI-proposed treatment plans or classification results replace clinicians’ judgments. It reports results from real-world trials and benchmark evaluations, but gives no details on the size of the trial datasets or the evaluation methods. Lunit released the model, and CEO Se Bumseok said the company would develop it as infrastructure for hospital-wide support. The reported hospital trial figures were 86.2% for multidisciplinary care, 98% agreement each for KTAS and ADR, and 85.6% for insurance review and claims.