LG AI Research has officially unveiled its next-generation AI foundation model, 'K-EXAONE,' at the '1st Public AI Foundation Model Presentation' hosted by the Ministry of Science and ICT. Developed as a frontier-class model boasting 236 billion parameters (236B), K-EXAONE signifies a substantial leap in AI technology, building upon the capabilities of the previous EXAONE 4.0.

The core methodology of K-EXAONE focuses on maximizing performance efficiency while simultaneously reducing computational demands and memory footprint. This is achieved through the integration of two key architectural innovations:

1. Mixture of Experts (MoE): K-EXAONE employs a proprietary Mixture of Experts (MoE) model structure. In an MoE architecture, the model consists of multiple "expert" sub-networks, and a "gating network" learns to selectively activate only a subset of these experts for each incoming input. This allows K-EXAONE to scale to a vast number of parameters, enabling sparse activation where only a fraction of the total parameters are engaged per inference, thereby significantly enhancing performance efficiency and scalability while reducing the computational cost compared to a dense model of similar parameter count.
2. Hybrid Attention Technology: Complementing the MoE structure, K-EXAONE incorporates a unique Hybrid Attention technology. This innovation specifically targets the self-attention mechanism, which is often a primary bottleneck in large transformer models due to its quadratic computational and memory complexity with respect to sequence length. The Hybrid Attention mechanism is designed to reduce memory requirements and computational operations by approximately 70%. While the article does not detail the exact nature of this hybrid approach, it typically involves combining different attention strategies (e.g., local, global, or sparse attention variants) or optimizing attention computations to achieve substantial efficiency gains, making the model more economically viable for deployment.

These architectural optimizations enable K-EXAONE to operate efficiently on A100-level GPUs without relying on high-cost, cutting-edge infrastructure, significantly lowering the barrier for startups and small and medium-sized enterprises to adopt frontier-class AI models.

In terms of performance, K-EXAONE has demonstrated superior capabilities against leading open-weight models. In a 13-benchmark evaluation, K-EXAONE achieved an average score of 72.03 points. This performance is notably higher than Alibaba's 'Qwen3 235B,' which scored 69.37 points (K-EXAONE achieved 104% of its performance), and OpenAI's 'GPT-OSS 120B,' which scored 69.79 points (K-EXAONE achieved 103% of its performance). According to Artificial Analysis's 'Intelligence Index,' GPT-OSS 120B and Qwen3 235B are ranked 6th and 7th globally among open-weight models, respectively. Based on these results, LG AI Research aims for K-EXAONE to enter the global TOP 5 among open-weight models.

Developed in just five months, K-EXAONE represents LG AI Research's accumulated expertise in AI foundation model development over the past five years. Looking ahead, LG AI Research plans to develop next-generation models with trillion-scale parameters, aiming to directly compete with top-tier global big tech models and solidify South Korea's position as a leading AI nation.