KULLM3

KULLM3 is a Korean-specialized large language model developed by the NLP&AI Lab at Korea University with the HIAI Research Center, unveiled in early 2024. It starts from Upstage's SOLAR-10.7B and applies instruction tuning on a large corpus of Korean instruction data, much of it generated specifically for the project.

The 10.7B-parameter model is chat-optimized and bilingual in Korean and English. The research team reported that its Korean responses matched or exceeded contemporary Korean models on fluency, coherence, accuracy, and completeness in GPT-4 Turbo-judged evaluations.

KULLM3 is one of the earlier serious attempts at a sovereign Korean instruction model, predating the wave of corporate open releases from SK Telecom, Naver, and Upstage. It remains a useful baseline for Korean NLP research; weights are publicly available but the exact license terms should be checked on the model card before commercial use.