NC AI, the artificial intelligence (AI) subsidiary of NCSOFT, said on the 31st that it will release its independent AI foundation model project, "VAETKI (Baeki)," as open source on the global open-source platform Hugging Face.
NC AI announced the completion of the first-stage development of Baeki and revealed industry-specific proof-of-concept results at the first presentation of the "independent AI foundation model project" held the previous day at the COEX Auditorium in Samseong-dong, Seoul. Baeki aims to go beyond a general-purpose large language model (LLM) to meet the specificity and security requirements of national key industry sites in manufacturing, defense, logistics, and content.
To develop industry-specialized AI models, NC AI built a "grand consortium" with 14 institutions across industry, academia, and research and 40 demand-side partners, including LOTTE Innovate, POSCO DX, MBC, KAIST, ETRI, and Korea University, completing an ecosystem that spans the full cycle from data acquisition to model development and proof-of-concept scaling.
The core of the Baeki model is industry suitability and cost efficiency. The company said that while global big tech firms are engrossed in competing to scale up model size, the NC AI consortium focused on optimizing efficiency so that corporations can actually adopt it. Baeki is a large model with more than 100 billion (100B) parameters, but by applying a mixture-of-experts (MoE) architecture, only 11B parameters are activated during inference. This maintains high performance while reducing operating expense.
It also reduces memory usage by up to 83% compared with existing models and increases computation speed, making it usable even in industrial sites lacking high-cost graphics processing unit (GPU) infrastructure.
NC AI said, "Baeki ranges from an ultra-large model (100B) to an on-premise lightweight model (sLLM) and a vision-language model (VLM) that processes visual information, forming a 'multi-scale' lineup that defense or semiconductor corporations, where security is paramount, can deploy on internal servers and use without worrying about data leaks."
The NC AI consortium is already delivering results by carrying out more than 28 diffusion projects in industrial sites using Baeki. In smart industry, it is working with consortium participant InterX to analyze process data at an auto parts corporation and optimize production lines, and it is pursuing collaborations with major domestic manufacturing corporations for industrial AI transformation (AX).
In defense and security, it is promoting defense-specialized AI cooperation with Army Headquarters to operate in a closed, secure on-premise environment. In distribution and logistics, it is discussing cooperation with consortium participant LOTTE Innovate to build a domain-ops platform.
Building on phase-one results, the NC AI consortium plans to accelerate technological advancement and global expansion through 2027. Next year, it plans to develop a 200B-class model (200 billion parameters) with global state-of-the-art (SOTA) performance and a large multimodal model (LMM) that comprehensively understands text, images, video, 3D, and sound. Based on this, the goal is to export "K-sovereign AI" in a package form to the Middle East and Southeast Asia by 2027.
Lee Yeon-su, CEO of NC AI, said, "NC AI's goal is not merely to build a high-performing model, but to help Korea's manufacturing, defense, and content industries take flight to the global top with the wings of AI," adding, "We will do our best to ensure that the AI technology NC AI has accumulated in games now runs real-world factories, protects national security, and becomes the core engine for spreading K-culture."