KT unveils Korean-specialized, GPT-4o-powered SOTA K

KT has launched SOTA K built on GPT-4o, a Korean specialized artificial intelligence (AI) model developed in a strategic partnership with U.S. tech giant Microsoft.

The new model will incorporate the Korean language and its social and cultural context on GPT-4o by collecting high-quality Korean-specific datasets and working with Microsoft to adapt the advanced AI model to the Korean context.

“SOTA K is an innovative model that combines world-class technology with Korean-specific AI specialization. The technical expertise we have gained through our collaboration with Microsoft will be applied across KT’s future AI developments,” Yoon Kyung-a, head of KT’s Agentic AI Lab, said Monday.

“This will not only revitalize the domestic AI ecosystem but also lead innovation in both the public and private sectors, boosting Korea’s overall AI competitiveness.”

SOTA K embodies four key principles of Korean AI — data sovereignty protection, cultural relevance, a two-track model strategy to choose between its own models and cooperative, open models, and reliable and responsible AI.

It can naturally process everything from Korean honorifics and dialects to industry-specific terminology in fields such as law, finance and history, making it highly practical for business and industry applications in Korea.

KT also established its own dataset to define and measure new benchmarks for Korean AI. In performance benchmarks that KT established with its own dataset, SOTA K outperformed GPT-4o in core areas including Korean language comprehension, reasoning, cultural understanding and specialized domain knowledge. In particular, it excelled in areas requiring advanced knowledge of Korea, such as history, law and language, even surpassing GPT-4o in mock versions of civil service and naturalization exams.

The company shared that real-world applications have already shown promise. Meritz Fire & Marine Insurance reported highly accurate policy summary reports and customer service scripts using SOTA K. And educational broadcaster EBS deployed the model to generate curriculum questions tailored to different grade levels, enhancing personalized learning.

At Yonsei University Health System, the model successfully converted mixed Korean-English medical data into clear and natural Korean that patients could easily understand, even handling details such as date formats in locally used ways. Korea Electric Power Corp. reported fast, clear and concise answers, and noted that system integration could expand its potential use significantly.

KT plans to first deploy SOTA K to its own business-to-consumer services to validate both proprietary and collaborative model lineups before rolling out services to partner companies.