Krafton has teamed up with SK Telecom to unveil three specialized 7B (7-billion parameter) inference language models.

The trio—dubbed OpenThinker2, OpenThinker3, and AceReason-Nemotron-1.1—focuses on cracking complex mathematical problems and code development. By deploying Krafton’s proprietary training techniques, these small language models have clocked in significant performance gains on the AIME 25 math reasoning benchmark. Given that mathematics demands both spatial awareness and logical deduction, it shares a clear technical lineage with the high-level reasoning required in gaming.

The collaboration saw Krafton and SKT divide and conquer: Krafton focused on refining training methodologies, while SKT took charge of infrastructure. Krafton developed a 'wrong-answer review' training technique, which identifies the weaknesses of existing models by comparing incorrect attempts against the correct solutions. Meanwhile, SKT handled the heavy lifting of data validation and infrastructure deployment, ensuring the models are both reliable and performant.

Kim Ji-won, head of the AI Model Lab at SK Telecom, noted that the joint venture serves as a foundation for their 'Sovereign AI' strategy. Kim Ji-won added that the team remains committed to securing AI technology sovereignty through continued close cooperation.

Lee Kang-wook, head of the Deep Learning Division at Krafton, stated that the partnership with SKT proves Krafton's capability to develop world-class large language models. Lee Kang-wook further mentioned that Krafton intends to leverage this proprietary AI prowess across various industries, including the gaming sector.

These language models are now live on the global open-source platform Hugging Face, where anyone can access and put them to use.