In a determined push to secure technological autonomy in an increasingly competitive global landscape, South Korea’s ambitious Sovereign AI Foundational Model project has officially advanced to its second phase, a critical milestone in the nation’s strategy to cultivate homegrown artificial intelligence. Spearheaded by the Ministry of Science and ICT (MSIT), this government-backed competition is designed to nurture the development of large language models (LLMs) powerful enough to rival international giants, thereby reducing dependency on foreign technologies. The transition has been anything but smooth, however, characterized by both remarkable technological progress from the selected finalists and a deep-seated controversy over the very definition of “sovereign” innovation. This debate has cast a long shadow over the project, raising fundamental questions about the balance between leveraging global open-source advancements and achieving true national technological independence, shaping the trajectory of the country’s AI future.
The Strategic Divergence of Finalists
The first phase of the rigorous competition concluded with the selection of three teams poised to lead the nation’s AI charge: LG AI Research, SK Telecom (SKT), and the dynamic startup Upstage. LG AI Research distinguished itself as the top performer, with its K-Exaone LLM earning the highest marks across all evaluation criteria for its exceptional accuracy, usability, and innovative design. Looking ahead to Phase Two, LG is channeling its efforts into developing highly specialized, industry-specific AI models. Its strategic focus is on creating tangible, practical applications for critical sectors such as power generation, the chemical industry, and biotechnology, aiming to build a comprehensive ecosystem that extends beyond model development to include industrial deployment and talent cultivation. In sharp contrast, the SKT consortium is charting a course toward multimodal capabilities, intending to move beyond text-only comprehension to a more holistic understanding of diverse data types. The initial step involves integrating image data to enable its A.X K1 model to perform tasks like recognizing and summarizing academic papers or business documents.
Building on the foundation of its A.X K1 model, which holds the title of South Korea’s first hyperscale AI with an impressive 519 billion parameters, SKT has laid out an ambitious roadmap. The consortium plans to significantly increase the volume of training data and expand the model’s linguistic capabilities to support five languages: Korean, English, Chinese, Japanese, and Spanish. The ultimate goal is the release of an improved successor, tentatively named A.X K2. Meanwhile, the AI startup Upstage is pursuing a different but equally vital strategy centered on strengthening its foundational technological capabilities and aggressively scaling its models. The company is actively recruiting leading AI scholars from prestigious global institutions, including Stanford and New York universities, to fortify its research and development team. Technologically, Upstage plans a phased approach to model growth, aiming to expand from approximately 100 billion parameters to 200 billion, and eventually to 300 billion, with the explicit goal of enhancing both the scale and performance of its AI. These distinct paths—LG’s industrial application, SKT’s multimodal expansion, and Upstage’s foundational scaling—collectively represent a multi-pronged national strategy to build a resilient and competitive AI ecosystem.
A Contentious Path and Its Fallout
The project’s initial phase was fraught with a widespread and contentious debate that struck at the core of its mission: the originality of the competing AI models. This controversy revolved around the fundamental question of what it means to develop a model “from scratch” in an era where leveraging external, open-source technology is common practice. Allegations were directed at several participants, creating a cloud of scrutiny. Upstage’s Solar-Open-100B model was accused of reusing inference code from a Chinese company’s AI, while SKT’s A.X K1 model faced similar scrutiny over alleged resemblances to the Chinese DeepSeek model. Both Upstage and SKT mounted strong defenses, arguing that the use of open-source inference code—which is responsible for running a pre-trained model on new inputs to generate outputs—is fundamentally distinct from the training code and the core model architecture. They maintained that this practice does not compromise the originality of their foundational models, a distinction that highlights the nuanced technical complexities in defining technological sovereignty.
The most significant casualty of this debate was the high-profile elimination of Naver Cloud, a participant widely considered a frontrunner in the competition. The expert evaluation panel concluded that Naver’s model failed to meet the requisite standard of originality, citing issues in both its technical implementation and its underlying policy framework. The company had acknowledged its use of external components from Alibaba’s open-source Qwen model, specifically its vision and audio encoders, as well as pre-trained parameters or “weights.” While Naver framed this as a strategic decision intended to enhance efficiency and ensure global compatibility rather than to undermine its technological independence, the panel’s decision sent a clear message about the project’s strict criteria. The fallout from this elimination underscored the high stakes involved and the government’s unwavering commitment to fostering genuinely homegrown innovation, even at the cost of excluding a major domestic player from its flagship AI initiative.
A Vacant Slot and Diverging Paths
Following the decision to advance only three of the originally planned four teams, the MSIT announced its intention to hold an additional selection process to fill the vacant fourth slot, a move aimed at restoring the competition to its intended scale. The ministry confirmed that all previously eliminated companies would be eligible to reapply, creating a “revival round” for those who did not make the initial cut. This decision reflects the government’s desire to keep the project robust and ensure that it moves forward with a full complement of participants, thereby maximizing the potential for innovation and collaboration. The open invitation was designed to provide a second chance for companies to align their models with the project’s stringent originality standards and re-enter the race to develop South Korea’s sovereign AI. However, the response from the disqualified tech giants suggested a fundamental divergence from the government’s vision.
The reaction from the major eliminated contenders was swift and decisive, signaling a potential schism in the national AI strategy. Both Naver Cloud and NC AI publicly stated that they had no intention of participating in the revival round, choosing instead to pursue their own independent AI development roadmaps outside the purview of the government-led project. This reluctance indicated that, while the sovereign AI initiative is prestigious, it was not considered the only viable path forward for these corporations, which will continue to invest in and develop their proprietary AI technologies. This development left the opportunity open for other initial participants, such as KT, Motif Technologies, Konan Technology, and a KAIST-led consortium, who reportedly considered reapplying for the final position. The episode concluded with a clear picture of a complex and evolving landscape where national ambition intersects with corporate strategy, setting a challenging but dynamic stage for the future of South Korea’s quest for technological independence.
