The Lee Jae-myung government will select five corporations to lead the development of a Korean-style large language model (LLM) as 'National Representatives' for artificial intelligence (AI). The plan is to develop a proprietary model that secures 95% of the performance compared to global AI models like OpenAI's ChatGPT and Google's Gemini. The goal is to build 'Everyone's AI' that all citizens can use based on this developed AI model. The AI industry anticipates that many domestic corporations that have developed their own LLM will apply.
On the 20th, the Ministry of Science and ICT announced that it will recruit a consortium of domestic AI corporations to participate in the 'Independent AI Foundation Model Project' until the 21st of next month. The large-scale AI model development project, originally promoted under the name 'World's Best LLM' during the previous Yoon Suk-yeol administration, has been renamed under the Lee Jae-myung government. The foundation model refers to a general-purpose model trained on large-scale data to be applicable across various fields.
The government plans to select up to five elite AI teams, supporting each team with a maximum of 1,000 graphics processing units (GPUs), annual data construction and processing expenses of 3 billion to 5 billion won, and resources needed for AI model development based on demand. The five elite teams will compete through a phased evaluation after being selected.
The AI foundation models to be developed include large language models (LLMs), multimodal models (LMMs), and large action models (LAMs). The Ministry of Science and ICT explained that during each evaluation phase, only the objectives will be set by the government, while the specific AI model development strategies and methodologies will be determined by the participating teams, applying the research methods of the Defense Advanced Research Projects Agency (DARPA).
According to industry sources, corporations that have developed their own LLMs among domestic firms are showing interest in the project. Companies considering participation include Naver, LG AI Research Institute, NC AI, ESTsoft, Konan Technology, Saltlux, and Upstage.
In the industry, there are strong indications that Naver, LG AI Research Institute, and NC AI will participate and be selected. Naver plans to accelerate the establishment of Sovereign AI based on its own LLM 'HyperCLOVA X.' Ha Jung-woo, appointed as the first head of AI policy for the Lee Jae-myung government, oversaw the development and operation of HyperCLOVA X during his tenure as head of the Naver Cloud AI Innovation Center. LG AI Research Institute's LLM 'ExaOne 3.5' was chosen in April as the only 'Notable AI Model' recognized by Stanford University in the U.S.
Among small and medium-sized enterprises, ESTsoft announced its intention to participate in the 'National Representative AI Project' upon officially launching its self-developed 'Allen LLM' on the 17th. The Allen LLM is an open-source based model specialized in report generation and inference using Retrieval-Augmented Generation (RAG). An ESTsoft official noted, 'We have constructed the Allen LLM from ultra-large models for data centers to lightweight models for on-device use,' adding that 'we have developed our own LLM optimized for AI search engines and established an independent technological ecosystem for agentic AI services.'
Konan Technology will also challenge the development business of the Korean-style independent AI model. Konan Technology unveiled its own developed Konan LLM in 2023 and released its inference model 'Konan LLM ENT-11' in May. Recently, it has accelerated the establishment of 'Korean-made AI infrastructure' in collaboration with the AI semiconductor startup Rebellion. The plan is to integrate Konan Technology's generative AI technology with Rebellion's neural processing unit (NPU) to build an independent AI ecosystem based on domestic technology and ultimately increase the self-reliance of sovereign AI technology.
A Konan Technology official stated, 'We have proven our domestic and international technical capabilities in the generative AI field, including being listed in Stanford University's ecosystem graph as a Korean-made foundation model,' and expressed an intention to prepare for participation in government projects.
Saltlux is also showing interest in this project. Saltlux announced the release of its latest 'Lucia3 LLM' specialized in language generation and understanding at the end of last month. AI startup Upstage has significantly improved the Korean performance of its next-generation LLM 'Solar Pro2.'
An industry official remarked, 'If selected as one of the five corporations, the project conditions will improve significantly, such as receiving support for GPUs that were hard to procure and being able to advance the AI model.' They added, 'Since this is a project deemed important by the government, even though competition will be fierce, many domestic technology corporations are expected to participate.'