#South Korea #Seoul #LG AI Research #AI Model #K-EXAONE

LG Introduces 'K-EXAONE': A New Era in AI

On January 12, 2026, LG AI Research, recognized as a prominent force in artificial intelligence within South Korea, unveiled its highly anticipated AI model named 'K-EXAONE'. With an impressive architecture boasting 236 billion parameters, it signifies a monumental leap in the nation's quest to establish itself as a formidable entity in the global AI landscape.

Exceptional Global Performance

K-EXAONE has already made waves on various international benchmarks, demonstrating outstanding performance that positions South Korea among the elite in AI technology. Notably, it secured the top spot in 10 out of 13 benchmark tests during the first-stage assessment of South Korea’s national artificial intelligence model project, with an impressive average score of 72 points, outpacing its competitors.

In terms of global rankings, K-EXAONE has achieved the remarkable position of 7th worldwide and ranks as the leading model in South Korea, according to the Intelligence Index curated by Artificial Analysis. This distinction marks K-EXAONE as the sole model from outside the US or China in the global Top 10 for open-weight AI models, highlighting South Korea's expanding influence in the AI field.

Innovative Technical Features

LG AI Research has invested over five years into the development of K-EXAONE, resulting in groundbreaking innovations that enhance both performance and cost-effectiveness. Key technological advancements include:

• - Hybrid Attention Mechanism: Combining sliding window attention for detailed analysis and global attention for broader context, this mechanism reduces memory and computation demands by 70% compared to its predecessor, EXAONE 4.0.
• - Advanced Tokenizer: With an expanded vocabulary of 150,000 words and optimized phrase grouping, K-EXAONE can handle documents 1.3 times longer than previous models, solidifying its processing capabilities.
• - Multi-Token Prediction (MTP): This feature allows K-EXAONE to predict multiple tokens simultaneously, resulting in a 150% increase in inference speed.

K-EXAONE operates with a Mixture of Experts (MoE) architecture, employing just 10% of its parameters (approximately 23 billion) actively while maintaining the capacity to process context lengths of 260,000 tokens—equivalent to over 400 A4 pages.

Open Accessibility and Innovation

In a move aimed at democratizing AI access, LG AI Research has designed K-EXAONE to run on A100-grade GPUs, making it viable even for companies with limited infrastructure. Shortly after its launch on the global open-source platform Hugging Face, K-EXAONE soared to #2 in global model trend rankings, indicating significant international interest.

Moreover, a detailed Technical Report outlining its architecture, training methods, and performance evaluations has been made available, reaffirming LG AI Research's commitment to transparency and open innovation in AI development.

Commitment to Safety and Ethics

Safety and trustworthiness are at the forefront of LG AI Research's priorities. The rigorous evaluation process includes a data compliance assessment, an internal AI ethics committee addressing global ethical standards, and a custom KGC-SAFETY index measuring safety and reliability in the Korean cultural context. K-EXAONE achieved an impressive score of 96.2 in this index, surpassing notable models like GPT-OSS 120B and Qwen3 235B.

Fostering Talent and Collaboration

In its endeavor to nurture future AI experts, LG AI Research is inviting over 50 domestic graduate students to participate in the K-EXAONE project as interns. Collaborations with leading academic institutions, including Seoul National University and KAIST, are also underway to further deepen the development of next-generation AI technologies.

Free API Access for Developers

To encourage innovation, LG AI Research is providing free API access to K-EXAONE until January 28 via the Friendli.ai platform. This initiative enables developers across the globe to create AI agents without the necessity for extensive coding knowledge or high-end infrastructure.

Statements from Leadership

Jungkyu Choi, Head of the Agentic AI Group at LG AI Research, stated, "K-EXAONE exemplifies how focused innovation can enable competition with global leaders despite resource limitations. We believe K-EXAONE will represent South Korea's ambitions in AI and contribute significantly to the global AI ecosystem."

Jinsik Lee, Head of the EXAONE Lab, noted: "In just our first version of K-EXAONE, developed under existing constraints, we laid the foundation for enhancing its performance towards a more advanced model. We are excited about the potential this model holds."

Kwang-mo Koo, Chairman and CEO of LG Corp., highlighted during his 2026 New Year’s message the importance of striving for innovation and exceeding previous standards, reinforcing the direction LG is headed toward in redefining global benchmarks in AI development.

In summary, LG AI Research positions K-EXAONE not just as another AI model but as a transformative player in the AI industry, dedicated to pushing technological boundaries and fostering an environment where innovation meets accessibility and ethical responsibility.