[News Terms] NCSoft Accelerates Development of Large Language Model (LLM)

LLaMa, HyperCLOVA X, KoGPT Released
'Barco (2.0)' Also Speaks English, Japanese, and Chinese

NCSoft is quietly accelerating the development of its artificial intelligence (AI) 'large language model (LLM)' VARCO, which was launched last year.

A 'large language model (LLM)' is a term that collectively refers to AI trained to understand and generate large volumes of human language. It is an expanded concept of a language model (LM) composed of artificial neural networks that possess countless numerical parameters called 'parameters'?numerical information added or modified when inputting commands to process data in the desired way by users.

Exterior view of NCSoft Pangyo R&D Center building.
[Photo by NCSoft]

LLMs generate appropriate responses by repeatedly training on vast amounts of text and grasping context. When given a sentence, they predict the next word after the previous one ends by understanding similarities between words and context formation to produce more accurate meanings.

Unlike traditional language models that learn fixed patterns, structures, and relationships within a given language scope, LLMs learn from large-scale language data to understand sentence structure, grammar, and meaning, enabling natural conversational interactions. They are also considered core technology for generative AI, which creates new content based on learned content patterns and inference results.

Representative examples include OpenAI's 'ChatGPT' and Meta's 'LLaMa.' In Korea, Naver introduced 'HyperCLOVA X,' optimized for the Korean language, Kakao's subsidiary Kakao Brain developed the Korean-specialized large-scale AI model 'KoGPT,' and LG AI Research Institute unveiled its own large-scale AI 'Exaone 2.0.'

According to the gaming industry on the 28th, NCSoft released the VARCO LLM 2.0 model at the end of last month through Amazon Web Services (AWS) Marketplace and the official GitHub page.

VARCO LLM 2.0 is the next-generation version of VARCO LLM, which NCSoft unveiled in August last year. It is divided into a small model with 7 billion parameters and a base model with 13 billion parameters. VARCO 2.0 is characterized by its proficiency not only in Korean and English, which VARCO previously supported, but also in Japanese and Chinese.

The amount of training data has also increased. VARCO LLM 2.0 used 4.5 times more pre-training data and 10 times more tuning data than the 1.0 model to enhance performance. Additionally, it learned from a self-built dataset considering AI ethics to provide user-friendly responses.

NCSoft plans to release a chatbot service language model called 'Dialogue' and a 'Story' model that can be used for actual game planning document creation based on VARCO LLM 2.0. It also plans to launch VARCO MLLM 1.0, a multimodal LLM capable of understanding not only text but also images and documents, in the second half of the year.