Media Group Saramgwa-Sup and MarkerAI consortium released the Korean LLM "Gukbap" model on HuggingFace in three versions: Gukbap-Mistral-7B, Gukbap-Qwen2.5-7B, and Gukbap-Gemma2-9B. The Gukbap-Gemma2-9B model scored 8.77 on Korean logical reasoning evaluation (Logickor), 4.5 on Korean cultural knowledge evaluation (K2-Eval), and 46.5 on Korean professional field evaluation (KMMLU) — achieving top scores among Korean-tuned versions of each global foundation model (Mistral 7B from France, Gemma-2 9B from US, Qwen-2.5 7B from China) through full fine-tuning specialized for Korean. Research lead Dr. Jeong Cheol-hyeon: "Many Korean LLMs are trained on datasets generated by top-tier models like GPT-4, which risks license violation lawsuits — we developed a method using only open-source teacher models without license restrictions to evolve seed data, proving top-tier performance is achievable." Media Group CEO Han Yun-gi: "Gukbap proves open-source LLMs alone can surpass large corporation LLMs — it will be a comfort to Korean developers researching with limited GPU resources and few options, and a good alternative for Korean government and enterprises concerned about security and dependence on foreign or large corporation LLMs."