Naver Cloud said on the 3rd that it will lead the government's "specialized artificial intelligence (AI) foundation model development" project in the cyber defense sector and build an ultra-large 700B (700 billion parameters) class model.
A total of 33 companies, schools, research institutes, and public organizations will join the project, including Naver Cloud, LG CNS, KEPCO KDN Co. (KDN), Korea Hydro & Nuclear Power Co. (KHNP), Financial Security Institute, Korea Institute of Science and Technology Information (KISTI), and LG Uplus. The models will be validated across seven fields: power, finance, science and technology, telecommunications, semiconductors, defense industry, and aerospace.
The consortium will develop two models in parallel, each specialized for defense and attack analysis, based on HyperCLOVA X and EXAONE. The plan is to combine the two models using a mixture-of-experts (MoE) approach to support the entire process from cyberthreat detection to response. For training, it will use about 830TB of security and industrial data secured by the participating institutions.
Computing resources were also deployed preemptively. In addition to 256 graphics processing units (GPUs) provided by the government, Naver Cloud used 4,000 B200s and LG used 256 H200s to conduct pretraining even before the project kickoff. The goal is to shorten development time and use the remaining time for performance improvements and field validation.
They will also develop technology to operate the model on closed networks with limited external internet access. The completed model will be released as open source for commercial use, and the plan includes free inspections for small and midsize businesses to detect vulnerabilities and cyberattacks, as well as support for security companies to commercialize AI services.