Naratmalsami Differs from US and China... Specialized AI Models for Hangul Emerging One After Another

AI Company Releases Korean LLM Open Source

In an ecosystem dominated by specific languages such as English and Chinese, AI models specialized in Korean are emerging one after another.

According to the industry on the 11th, More, an AI infrastructure solution company, has open-sourced its self-developed large Korean language model (LLM) 'Motif' on the global AI platform 'Hugging Face'.

Motif utilized not only texts collected from websites but also specialized documents such as domestic patents and research reports as training data. As of the 3rd of this month, Motif scored 64.74 on 'KMMLU', the Korean AI performance evaluation benchmark, showing higher performance than global big tech companies OpenAI and Meta.

AI specialist company Dinotics also announced that it has open-sourced its self-developed LLM foundation model 'DNA' on Hugging Face and will begin beta testing of its generative AI assistant.

In the KMMLU benchmark, which evaluates humanities, social sciences, and science and technology in both Korean and English, DNA recorded an average score of 53.26. The company stated that this figure surpasses LG's 'Exaone 3.5' and NCSoft's 'Barco'.

The intensifying competition in Korean language-specialized AI models is evaluated to stem from the limitations of global big tech AI models. AI models undergo pre-training, which learns basic patterns before full training, and fine-tuning, which optimizes AI for specific fields. The data used in these processes is mostly known to be based on English and Chinese. This raises concerns that developers and companies using open-source AI models may overlook translation errors and cultural differences that can arise from models centered on English and Chinese.

Naver’s sovereign AI strategy is also aimed at resolving these issues. Naver is promoting cooperation with countries or companies that introduce AI models reflecting their own culture, such as Saudi Arabia.

There is also a view that by open-sourcing AI models and allowing developers worldwide to test and utilize Korean language models, it can help foster a Korean language-specialized AI ecosystem in the long term. Beyond natural language, Korean AI models specialized in image recognition and professional fields such as law and medicine have already appeared. The Korean language-specialized AI model market is expected to become more active.

NCSoft has released 'Barco Vision', a small-to-medium-sized open-source visual language model (VLM) specialized in Korean language processing, and OpenAI announced last month, through a memorandum of understanding (MOU) with Korea Development Bank for AI ecosystem development, its plan to develop AI models tailored to Korean contexts. More plans to launch and open-source 'Motif Vision', which generates images from text input, this month.

Professor Choi Kyung-jin of Gachon University Law School said, "If AI that does not understand Korean contexts generates images, it might depict children playing Korean traditional games wearing Chinese costumes," adding, "The core of AI level is diverse adaptability; it needs to learn large amounts of data from various regions to provide more accurate and diverse answers."