Two AI-focused foundation models for the medical and bio sectors being developed in South Korea have confirmed global-level performance and will enter phase two development.

The Ministry of Science and ICT said on Monday it announced interim performance evaluation results for the "AI-focused foundation model" project with the National IT Industry Promotion Agency (NIPA) and the Telecommunications Technology Association (TTA).

A Lunit consortium and a KAIST consortium have been independently developing AI-focused foundation models for medical science and bio fields from scratch since November last year. In the interim evaluation, both consortia scored above 80 points, meeting the phase two support standard of at least 70 points. The ministry plans to continue providing each consortium with 256 Nvidia B200 GPUs through early September.

The Lunit consortium's 16-billion-parameter medical science foundation model uses a mixture-of-experts (MOE) approach. It delivered strong results in medical paper-based question-and-answer accuracy, consistency between sources and evidence, and scientific code writing and analysis evaluations, compared with very large general-purpose models in the 100-billion to 1-trillion-parameter range such as Claude 3.5 Sonnet by Anthropic.

Based on the model, the Lunit consortium has also built a Clinical Decision Support Agent System (CDSS). In tests applying it to a real medical data environment at the National Health Insurance Service Ilsan Hospital in February and March, it showed high accuracy in five-level emergency room patient triage and recorded a 94.0 percent match rate for diagnosis names.

The KAIST consortium's 2-billion-parameter bio foundation model, K-Fold, was assessed as close to Google AlphaFold3 in accuracy for predicting the three-dimensional structure of protein complexes, while being up to more than 30 times faster in prediction speed. It reduced average prediction time to within 1 minute, compared with AlphaFold3's structure prediction time of about 30 minutes. KAIST newly applied a structure prediction method based on physical and chemical interactions, rather than existing methods, to improve prediction accuracy even for new drug complexes with sparse data.

The models developed by the two consortia will be released in early April on Hugging Face as open source under the Apache 2.0 license. The Lunit consortium plans in phase two to scale the model to up to 32 billion parameters and conduct field demonstrations in July and August at nine hospitals and at SK Biopharm, among others. The KAIST consortium plans to scale to a 7-billion-parameter model and pursue entry into overseas markets by deploying it on global drugmaker Merck's AI drug development cloud service.

Choi Dong-won (최동원), director general for AI Infrastructure Policy at the ministry, said, "Even in a period of about 5 months, we have built a technological foundation that can work in the global market." He added, "We will not spare policy support so that industrial application potential in high value-added industries such as diagnosis and treatment and new drug development can lead to actual commercialisation."