A New Korean Open-Source AI Model: Unveiling Its Comprehensive Components

The Korea Advanced Institute of Science and Technology (KAIST) has launched KORMo-10B, a groundbreaking open-source Korean-English language model comprising 10.8 billion parameters. Spearheaded by Professors Lim Kyung-tae and Oh Hye-yeon, KORMo stands out for its complete transparency—unveiling all components, including training data, code, and intermediate models—available on the Hugging Face platform. This initiative aims to provide 95% of the performance of state-of-the-art (SOTA) models, bridging a significant gap in Korean language AI capabilities, often lacking in existing global models. The development process utilized a majority of synthetic data to enhance Korean representation. KORMo serves as a versatile tool for domestic companies, allowing users to customize it by integrating their data. This release marks a milestone, being the first fully open-source large language model for non-English languages, promoting accessibility and innovation in AI for public interest.

Source link