South Korea Selects 5 Elite Teams for National AI Project

Naver Cloud, Upstage, SK Telecom, NC AI, and LG AI Research have been selected as the five elite teams to participate in the government’s "Independent AI Foundation Model" project."

The Ministry of Science and ICT announced on Aug. 4 that it has selected five teams as finalists for the national AI champion selection project through document evaluations and presentations. The ministry initially selected 10 organizations through the first document evaluation, including Naver Cloud, Motif Technologies, Upstage, SK Telecom, NC AI, LG AI Research, Kakao, KT, Konan Technology, and KAIST. Subsequently, the government finalized five elite teams through presentation evaluations. The government assessed the development goals, strategies, ripple effects, and contribution plans of the participating elite teams in this presentation evaluation. The government plans to gradually reduce the number of AI models eligible for government support by one through phase evaluations every six months, following the selection of five teams. The selected companies and institutions will receive support worth 200 billion won, including high-performance graphics processing units (GPUs), data, and labor costs.

Kakao, KT, Korea Advanced Institute of Science and Technology (KAIST), Konan Technology, and Motif Technologies were eliminated.

Naver Cloud, selected as one of the five finalists, has formed a consortium with Naver, Silicon Valley video multimodal AI startup Twelve Labs, and leading domestic universities and research institutions including KAIST, Seoul National University, POSTECH, Korea University, and Hanyang University.

Naver Cloud possesses AI full-stack capabilities, which enable it to perform the entire process from AI technology development to service implementation with its own technology. From the basic stage, it has directly built and operated everything from its self-developed large-scale language model "HyperCLOVA X" to large-scale data centers, cloud infrastructure, AI platforms, applications, and user services, accumulating experience in actual large-scale service applications.

In this project, Naver will combine its language and voice-based multimodal technology with Twelve Labs’ video AI technology to develop an "Omni Foundation Model" that comprehensively understands and generates heterogeneous data such as text, images, audio, and video. Through this, they plan to implement a complete multimodal AI that can be used not only in domestic industrial and public sectors but also in the global market.

The Naver Cloud consortium plans to provide AI services that all citizens can easily experience based on this model, and support anyone to develop, register, and distribute AI agents through an AI agent marketplace. They will also pursue the commercialization of industry-specific models and solutions.

Furthermore, based on their experience in building sovereign AI in Saudi Arabia, Thailand, and Japan, they will establish a K-AI global export model and contribute to the expansion of the domestic AI ecosystem through open-sourcing of lightweight and inference-specialized models.

Sung Nak-ho, CTO of HyperCLOVA at Naver Cloud, said, “We believe that AI should create value in actual services and industrial fields beyond just being well-made. Through this project, we will implement a national representative AI model that can be used anywhere at home and abroad, and create K-AI that works on the global stage.”

Meanwhile, the SK Telecom consortium, which includes game companies Krafton, 42dot, Rebellions, Liner, and SelectStar, was launched with the goal of implementing a full-stack AI based on independent technology from semiconductors to models, data, and services, and opening the developed models as open source to various companies in the domestic AI ecosystem. The consortium provides AI agent services familiar to customers, including "A.," the top-level domestic communication agent with 10 million users, and Liner, which specializes in research and information search based on accurate information provision and source selection capabilities.

42dot is pursuing global expansion by applying the in-vehicle conversational agent "Gleo AI" to Hyundai Motor and Kia. Krafton has demonstrated its technological prowess by revealing co-playable characters (CPC) that interact like humans based on SLM co-developed with NVIDIA in the next-generation global game ‘inZOI,’ as well as "Orak," a benchmark that evaluates AI agents’ game-playing abilities. 42dot and Krafton will take on key roles in on-device model development and multimodal design in the consortium, respectively, to enhance the scalability and practicality of the independent AI model. The field data and technical know-how accumulated through this project will be converted into high-quality training datasets through SelectStar’s data processing platform "Cashmission," which has a scale of 240,000 people. SelectStar evaluates AI model stability based on "Datumo Eval," Korea’s first LLM reliability verification solution, and experience in operating domestic and international generative AI red team challenges.

The SK Telecom consortium will be led by Kim Tae-yoon, SK Telecom’s foundation model manager, who has been leading the development of the "A. X" model. Research teams from Seoul National University (Profs. Kim Gun-hee, Yoon Sung-ro, Hwang Seung-won, Do Jae-young), KAIST (Prof. Lee Ki-min), and the University of Wisconsin-Madison (Profs. Lee Kang-wook, Dimitris Papailiopoulos) will conduct original technology research. Over 80% of the total project personnel from the consortium’s main and participating companies are master’s and doctoral degree holders, with about 120 AI experts leading the original technology research. The consortium also boasts the highest level of competitiveness in Korea in terms of the number of papers (800), patents (736), and open sources (270) held by the research team.

The industry’s expectations for the consortium’s research team are high, with about 20 companies and institutions, including SK Group companies such as SK Hynix and academic institutions like the Korea Foundation for Advanced Studies, having submitted letters of intent to use the model.

Through this Independent AI Foundation Model project, SK Telecom plans to challenge the development of a large-scale AI model that surpasses the scale of existing domestic LLMs. It will also implement high-performance, high-efficiency AI services through optimization of domestic NPU (Neural Processing Unit) utilization technology possessed by Rebellion.

Kim Ji-won, SK Telecom’s AI Model Lab head, stated, “We will realize the highest level of Korean-style independent AI foundation model for AI in citizens’ daily lives with the prepared technological prowess and execution ability of industry-leading companies.”

NC AI, the artificial intelligence (AI) specialist subsidiary of NCSoft, has been selected as a finalist for the government’s Independent AI Foundation Model development project. NC AI will be able to use designations such as “K-AI Model” and “K-AI Company.” The company will receive support for computing resources necessary for AI model advancement, data for model training, and access to a pool of development talent.

The NC AI consortium plans to develop a foundation model with performance at least 95% comparable to global AI models released within the past six months. NC AI intends to support industrial AI transformation through its domain ops platform business and business-to-business (B2B) system integration (SI) services. Furthermore, the company aims to contribute to the AI ecosystem by connecting multimodal cognitive and generative technologies with public government services.

NC AI emphasized that it will develop the best AI model in collaboration with 13 domestic and international industrial and research institutions in the government AI foundation model development project consortium. These include Korea University, Lotte Innovate, MediaZen, Seoul National University, Yonsei University, AIWorks, InterX, POSCO DX, Korea Advanced Institute of Science and Technology (KAIST), MBC, NHN, HL Robotics, and Electronics and Telecommunications Research Institute (ETRI).

NCSoft established an independent AI organization within the company in 2011 and created a natural language processing (NLP) research team under its AI Lab in 2015, becoming the first domestic game company to research AI capable of understanding and producing Korean sentences. In July 2023, it introduced “VARCO,” the first generative AI brand in the gaming industry. NCSoft unveiled the "VARCO 2.0" large language model (LLM) with 7 billion and 13 billion parameters in May of the following year, and in December, launched "VARCO Vision," a vision-language model (VLM) specialized in Korean language processing.

The AI business has gained further momentum since spinning off as NC AI in February this year. NC AI’s strategy is to expand its AI technology, which has been primarily focused on application in the gaming sector, to fashion, media, and content fields, and to create a revenue model based on Software as a Service (SaaS).

NC AI has recently unveiled a series of AI models. On July 30, it open-sourced "VARCO Vision 2.0 1.7B," an ultra-lightweight multimodal vision-language model (VLM). According to NC AI, this model can simultaneously analyze multiple images to process complex documents, tables, and charts. As a lightweight model with 1.7 billion parameters, it can operate in on-device environments such as personal computers or smartphones. NC AI explained that the performance of this model is comparable to open-source multimodal models like InternVL3 2B and Obis2 2B, based on major benchmark scores such as SeedBenchIMG and LavaBench.