[Edaily Kim Hyun-ah Reporter Yoon Jeong-hoon] Starting in August, South Korea’s artificial intelligence (AI) industry is set to face a major test of its “indigenous AI” capabilities. Four companies participating in the government’s competition for indigenous AI foundation models—Upstage, Motif Technologies, SKTelecom(017670), and LG Corp.(003550) AI Research Institute—have begun seeking validation in the global market by successively releasing their own ultra-large AI models on the global AI platform “Hugging Face.”

The focus of the AI competition is shifting from mere model scale to efficiency and usability. Korean companies are highlighting Korean language processing capabilities, inference efficiency, and applicability in corporate settings as their differentiating factors.

With the government-led “Second Evaluation of Independent AI Foundation Models”—in which one company will be eliminated—as well as the “AI for All” and “Security-Specialized AI” projects all concentrated in August, the K-AI competition is expanding beyond model development to include competition in services and industrial applications.

Upstage’s “Solar Open 2”… Korean Language Capabilities Draw Attention

One of the models attracting significant attention on Hugging Face is Upstage’s “Solar Open 2.”

With 250 billion (250B) parameters, Solar Open 2 has recorded over 13,000 downloads since its release, showing rapid adoption among domestic proprietary AI models. It is also being evaluated against global models on aib.vote, a domestic AI model comparison platform.

Its strengths lie in its ability to express Korean naturally and process long contexts. It supports a context window of up to 1 million tokens, focusing on enhancing the analysis of long documents and its applicability to corporate workflows.

However, since solving complex problems requires additional inference operations, striking a balance between response speed and cost-efficiency remains a key challenge.

Motif Takes on the Challenge with Its Proprietary AI Design

Motif Technologies

,

a subsidiary of AI infrastructure company More, showcased its technical capabilities by unveiling “Motif-3-Beta.”

With 314 billion (314B) parameters, Motif-3-Beta is notable for being developed based on a proprietary architecture rather than simply being a modified version of an existing model.

Its core technology is the sparse routing method. This approach reduces computational load and improves efficiency by selectively activating only the necessary networks within a Mixed-Expert (MoE) architecture.

However, as commercial use is restricted in the beta version, discussions regarding its practicality for developers are ongoing. Motif stated, “This is a temporary condition related to the Sovereign AI project, and we plan to lift the restriction in the final public release.”

SKTelecom’s ‘A.X K2’… Even Ultra-Large Models Are Competing on Efficiency

SKTelecom has unveiled “A.X K2,” a Mixed Expert (MoE) model with 688 billion (688B) parameters.

MoE is a technique that selects only the parts necessary for a given query from among multiple expert networks; while the overall model size is large, this approach reduces the actual computational load.

A.X K2 has introduced hybrid inference technology that applies “Think Mode” to complex problems and uses a standard response method for simple queries.

LG Corp.’s ‘K-EXAONE 2.0’… Targeting the Enterprise AI Market

LG Corp. AI Research’s ‘K-EXAONE

2.

0,’ with 750 billion (750B) parameters, is the largest model among those unveiled this time.

Having taken first place in the initial evaluation, LG Corp. has positioned safety and reliability as its core competitive advantages, targeting the enterprise AI market. The company has focused on enhancing the stability and security required for long-text document understanding, search, and enterprise work environments.

As LG Corp., Upstage, SKTelecom, and Motif—which participated in the government’s independent AI foundation model competition—enter the global open competition, technological competition among domestic AI companies is also heating up.

Although not a participant in the government’s independent AI initiative, Naver is also joining the competition by pushing forward with the development of the next-generation HyperClova X, which will feature 500 billion (500B) parameters.

“In the Era of Agent-Based AI, Korean Language Efficiency Is a Competitive Edge”

The decisive factors in the AI competition are shifting from model size to actual task performance and cost efficiency.

In particular, in the era of agentic AI—where AI independently devises plans and utilizes tools—the efficiency of Korean language processing is emerging as a key competitive factor.

Lee Dong-soo, CEO of AI infrastructure platform company a2sys, stated, “In the era of traditional chatbots, the difference in token efficiency between Korean and English was not significantly noticeable, but in an agentic AI environment, this difference could widen into a gap in cost and performance.”

A token is the basic unit by which AI processes text; even for sentences with the same meaning, the number of tokens required varies depending on the language and tokenizer design. Since agentic AI repeatedly performs searches, verifications, and tool executions, even a small difference in efficiency can impact a service’s competitiveness.

CEO Lee emphasized, “A competitive Korean-style AI service can only be created by combining a Korean-specialized tokenizer with high-quality data, industry-specific knowledge data, and the ability to utilize tools.”

Experts point out that it will be difficult to secure global competitiveness if the race to develop proprietary AI is limited solely to the development of super-large models. They explain that competitiveness is needed across the entire AI ecosystem—including not only large language models (LLMs) but also lightweight models, embeddings, speech and document AI, search technology, and safety evaluation systems.

The competition among four domestic independent AI models released on Hugging Face serves as the first test for validation in the global market. With the government-led announcement of independent AI models, the “AI for All” initiative, and security-specialized AI projects following suit, the K-AI competition is expected to expand beyond model performance to include competition in actual service implementation and industrial applications.

The decisive factor will likely be not the size of the model, but rather securing “real-world competitiveness”—the ability to demonstrate cost-effective performance in actual industrial settings.