The homegrown artificial intelligence (AI) foundational model, K-Exaone, developed by LG AI Research, has entered the global top 10, ranking seventh. This achievement marks K-Exaone as the only Korean model in a ranking largely populated by models from the United States and China.
In a statement, LG highlighted that its latest AI model excelled in a government-led competition, outperforming five teams and achieving top results in 10 of 13 benchmark tests with an average score of 72. The Intelligence Index compiled by Artificial Analysis placed the model seventh internationally, with China dominating the list with six models and the U.S. featuring three. Z.AI’s GLM-4.7 secured the top position.
LG made its foundational model available as an open-weight on Hugging Face, where it quickly climbed to second place on the platform’s global model trend chart. This surge indicates significant interest from international stakeholders. Furthermore, LG announced plans to offer free API access to K-Exaone until January 28, allowing developers and companies to utilize the model without incurring costs during the introductory phase.
Epoch AI, a U.S.-based nonprofit, has also acknowledged the model, adding it to its list of notable AI models. With five models included, LG AI Research now holds the distinction of having the most entries from a Korean company. “We established the development plan according to the time and infrastructure we were given, and we developed the first-phase K-Exaone using about half the data we have,” said Lee Jin-sik, head of Exaone Lab at LG AI Research.
According to LG, the K-Exaone model is the result of five years of in-house research and signifies Korea’s commitment to competing in the global frontier-class AI systems landscape. The company emphasized that rather than solely relying on scale, it reengineered its architecture to enhance performance while reducing training and operational costs. The model features a mixture-of-experts (MoE) architecture with 236 billion parameters, activating around 23 billion parameters per inference.
The K-Exaone model distinguishes itself through its core technology, hybrid attention, which improves its ability to prioritize crucial information during data processing while cutting down computational load by 70% compared to prior models. Additionally, it has upgraded its tokenizer, expanding its training vocabulary to 150,000 words, and optimizing frequently used word combinations to enhance document processing speed by 1.3 times.
Multi-token prediction has further accelerated inference speed by 150%, improving overall efficiency. “K-Exaone is designed to maximize efficiency while reducing costs, allowing it to run on A100-class GPUs rather than requiring the most expensive infrastructure,” an LG AI Research official stated. This accessibility promotes the development of frontier-level AI for companies with limited computing capabilities, thereby broadening Korea’s AI ecosystem.
Beyond mere memorization, K-Exaone is designed to enhance its reasoning and problem-solving skills. During its pre-training stage, the model was exposed to thinking trajectory data, which illustrates how problems are solved rather than just providing final answers. LG has prioritized safety and compliance, conducting thorough data compliance reviews across training datasets to eliminate materials with potential copyright issues.
To ensure ethical standards, LG operates an internal AI ethics committee that assesses risks across four categories: social safety, Korea-specific considerations, future risks, and universal human values. Under its KGC-Safety benchmark, K-Exaone scored an average of 97.38 across these categories, outperforming OpenAI’s GPT-OSS-120B model and Alibaba’s Qwen-3-235B model. This performance not only underscores the technical capabilities of K-Exaone but also positions LG AI Research as a significant player in the rapidly evolving AI landscape.
See also
LG AI Research Institute Achieves Top Scores in National AI Model Benchmark Tests
AI Expert Nathan Lambert Reveals Multi-Model Strategy for Optimal Performance
Arizona Lawmakers Advance Psychedelic Legislation Amid Rising AI Psychosis Concerns
Deep Learning Study Reveals Personalized Entrepreneurship Education Model Enhancements
Owkin Launches AI Agents for Drug Discovery, Enhancing Clinical Success Rates with Real Patient Data
