Krafton and SKT Co-Develop AI Models, Released as Open Source

Three Small Language Models Specialized in Math Problems and Code Development Unveiled
New Training Techniques Developed to Enhance LLM Mathematical Reasoning
"Contributing to Securing AI Technology Sovereignty"

Krafton announced on the 28th that it has unveiled three inference-specialized artificial intelligence (AI) language models with 7 billion (7B) parameters, co-developed with SK Telecom. The released models are OpenThinker2, OpenThinker3, and AceReason3.

These newly released models are small language models (SLMs) specialized in solving mathematical problems and code development, and Krafton applied its proprietary training techniques. Based on these techniques, the models achieved improved performance on the mathematical reasoning benchmark 'AIME 25'.

According to both companies, mathematics requires spatial perception and logical reasoning abilities, which are technically related to high-level inference fields such as gaming.

Krafton and SKT jointly developed the language models, with Krafton responsible for improving training techniques and SKT for building the necessary infrastructure.

Specifically, Krafton independently developed an error review training method by analyzing the weaknesses of existing models. This method involves finding the correct answers to previously incorrect problems and comparing them to the wrong answers, thereby improving both reasoning accuracy and efficiency. SKT was in charge of data validation and building the infrastructure for model training, ensuring the quality and stability of the models.

Kim Ji-won, head of AI Model Lab at SK Telecom, said, "With the technologies 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."

Lee Kangwook, head of Deep Learning Division at Krafton, stated, "Through this collaboration with SKT, we have confirmed our capability to develop large-scale language models at a global level," and added, "Krafton plans to apply its proprietary AI technology to various industries, including gaming."

Krafton plans to utilize the training techniques applied to these language models to further advance game-specific AI applications such as gameplay analysis and strategic decision-making. At the same time, the company intends to develop LLMs of various sizes.

Meanwhile, the newly released language models are available to everyone via the global AI open-source platform 'Hugging Face'.