Krafton Unveils Language Model Co-developed with SKT... Accelerating Development of Korean LLM

[Financial News] Krafton has unveiled three inference-specialized language models with a scale of 7B (7 billion parameters) co-developed with SK Telecom.
According to Krafton on the 28th, the models released this time are small language models specialized in solving math problems and code development, applying a learning technique independently developed by Krafton.
Based on this technique, the model demonstrated its effectiveness by recording a significant performance improvement in the math reasoning benchmark AIME 25. Mathematics requires spatial perception and logical reasoning capabilities, which are technically closely related to high-level reasoning fields, including games. Thanks to these capabilities, Krafton is also anticipating the possibility of expanding game-centered AI technology based on this model.
Krafton and SKT co-developed the language model, contributing to the enhancement of the model's quality and performance through improvements in learning techniques and infrastructure building, respectively. This collaboration is evaluated as a case that demonstrates the capability to develop domain-specialized AI models.
Krafton independently developed a learning technique that reviews incorrect answers by analyzing the weaknesses of existing models and improving them. This technique learns by finding the correct answers to wrong questions and comparing them with incorrect answers, ensuring inference accuracy and efficiency simultaneously. SKT contributed to securing the model's quality and stability by being responsible for data validation and infrastructure building for model learning.
Kim Ji-won, head of SK Telecom AI Model Lab, said, "With the technology of both companies, we have laid the foundation for realizing the sovereign AI strategy by developing high-performance language models," adding, "We will continue to contribute to securing AI technology sovereignty through close cooperation."
Kang-wook Lee, head of Krafton Deep Learning Division, said, "Through this collaboration with SKT, we were able to confirm the capability to develop large-scale language models at a global level," adding, "Krafton plans to apply this based on its proprietary AI technology to various industrial fields, including games."
Krafton plans to utilize the learning technique applied to this language model in enhancing game-specialized AI application technologies such as game play analysis and strategy judgment. In particular, by developing LLMs of various scales, they plan to contribute to securing the independence and leadership of the Korean AI technology ecosystem.


yjjoe@fnnews.com Yoon-joo Cho Reporter