The South Korean government is accelerating the development of globally competitive, specialized AI foundation models in the medical and bio sectors. The ‘AI Specialized Foundation Model’ project aims to strengthen Korea’s AI ecosystem by creating unique models tailored to local needs, in addition to general-purpose AI. The initiative is led by the Ministry of Science and ICT, the National IT Industry Promotion Agency, and the Korea Information and Communication Technology Association. The project was launched to secure Korea’s position in high-value industries such as diagnostics and drug development.
The Lunit consortium is developing the world’s first end-to-end medical science AI foundation model, while the KAIST consortium is focused on a bio foundation model called K-Fold, which predicts protein complex structures faster and more accurately than Google AlphaFold3. Both consortia passed mid-term performance evaluations, demonstrating superior results in key benchmarks. Lunit’s 16B model outperformed larger global models in medical Q&A accuracy and clinical decision support, while KAIST’s 2B K-Fold model achieved up to 30 times faster predictions with high accuracy for protein and molecular complexes.
The project began in November 2023, with both consortia receiving B200 GPUs (256 units each) from the Ministry of Science and ICT. After a successful mid-term evaluation, both teams scored over 80 points, exceeding the 70-point threshold for continued support. The second phase, from April 1 to September 9, 2024, will see ongoing GPU support and further development. Models are scheduled to be released as open source on Hugging Face in early July, following verification by the Korea Information and Communication Technology Association.
Frequently asked questions include: What are the practical impacts of these models? Lunit’s model has been tested in clinical environments, showing high accuracy in emergency patient triage and drug reaction analysis. KAIST’s K-Fold model offers rapid and precise protein structure prediction, beneficial for drug development. When will the models be available? Both models are expected to be open-sourced in early July 2024. How is the government supporting commercialization? The Ministry of Science and ICT has committed ongoing policy and resource support to ensure these innovations translate into real-world applications.
South Korea’s targeted approach to developing specialized AI foundation models in medical and bio fields is producing tangible results, with both Lunit and KAIST consortia achieving high performance in global benchmarks. The government’s sustained support, including GPU resources and policy backing, ensures continued progress and practical application. The open-source release of these models will facilitate wider adoption and innovation, particularly in diagnostics and drug development. The rapid advancement within a short timeframe demonstrates Korea’s commitment to global competitiveness and industry transformation.