LG AI Research has unveiled its world-class artificial intelligence (AI) model, K-EXAONE, joining the global competition for AI model supremacy.
According to industry reports on January 11, K-EXAONE topped 10 out of 13 benchmark tests, which were the primary evaluation criteria for independent AI foundation model projects. It achieved an impressive average score of 72, outperforming models developed by five elite teams.
In the Intelligence Index assessment by Artificial Analysis, a global AI performance evaluation organization, K-EXAONE scored 32 points. This ranks it 7th globally and 1st in South Korea among open-weight models. With the global Top 10 open-weight models dominated by China (6) and the U.S. (3), K-EXAONE stands as the sole Korean representative, upholding national pride in AI innovation.
K-EXAONE has garnered significant international attention, quickly rising to second place in global model trend rankings after its release on Hugging Face, a leading open-source AI platform. The U.S. nonprofit AI research organization Epoch AI has recognized five LG models in its Notable AI Models list, including EXAONE 3.5 (2024), EXAONE Deep, EXAONE Path 2.0, and EXAONE 4.0.
Stanford University’s annual AI report uses the Notable AI Models list to assess national and corporate competitiveness in the field.
LG’s Proprietary Hybrid Attention Technology Drives K-EXAONE’s Success
K-EXAONE represents the culmination of LG AI Research’s five-year journey in AI foundation model development. The model incorporates the advanced Hybrid Attention technology, which acts as the AI’s cognitive center, determining critical information focus when processing vast datasets.
LG AI Research has refined the Hybrid Attention by combining Sliding Window Attention for specific information focus with Global Attention for overall context comprehension. This innovation has slashed memory and computational requirements by 70% compared to EXAONE 4.0.
The team also enhanced the tokenizer, crucial for AI language proficiency. By expanding the vocabulary to 150,000 words and implementing efficient word-pair combinations, K-EXAONE can now process documents 1.3 times longer than its predecessors.
The introduction of Multi-Token Prediction (MTP) has boosted inference speed by 150% over earlier models.
K-EXAONE boasts 236 billion parameters, with 23 billion active parameters in an expert mixture model. It can process an unprecedented 260,000 tokens at once, equivalent to over 400 A4 pages.
The model’s learning process goes beyond mere data memorization, incorporating logical problem-solving skills. It uses innovative algorithms like Agapo for reinforcement learning from errors and Grouper for adopting human-preferred natural language patterns.
Ensuring Safe and Trustworthy AI: Transparent Technology and Talent Development
LG AI Research maintains strict AI ethics standards, conducting thorough data compliance assessments and excluding copyrighted material. An internal AI ethics committee oversees risk classification based on universal values, social safety, Korean cultural nuances, and future risk mitigation.
In the KGC-SAFETY index, developed to evaluate Korean-specific factors, K-EXAONE scored an impressive 97.83 across four categories. This surpasses global competitors like GPT-OSS 120B (92.48) and Alibaba’s Qwen3 235B (66.15) in safety and reliability.
LG AI Research has publicly released a comprehensive technical report on Hugging Face, detailing K-EXAONE’s architecture, training methodologies, and performance metrics.
The company is actively recruiting interns for the K-EXAONE project, aiming to provide over 50 domestic graduate students with hands-on experience in cutting-edge AI foundation model development.
Free Access to K-EXAONE Until January 28
LG AI Research is offering free access to the K-EXAONE API until January 28. This initiative allows AI enthusiasts to develop and deploy custom AI agents without the need for high-end infrastructure or specialized coding skills.
Choi Jung-kyu, head of LG AI Research’s Agentic AI Group, stated that K-EXAONE showcases our ability to compete with global AI giants through innovative technology, despite resource constraints. It is committed to advancing South Korea’s AI capabilities and contributing to the global AI ecosystem.