LG AI Research on Tuesday unveiled K-Exaone, a 236-billion-parameter artificial intelligence model that it said outperforms leading global AI systems by more than 100 percent, marking a major milestone in Korea’s push to develop homegrown foundation models.
The model was introduced at the first sovereign AI foundation model presentation hosted by the government at COEX in southern Seoul. LG said K-Exaone outperformed comparable large-scale models from global tech leaders, including Alibaba’s Qwen3 235B and OpenAI’s GPT-OSS 120B.
In benchmark testing, K-Exaone posted an average score of 72.03, surpassing Qwen3’s 69.37 and GPT-OSS’s 69.79 — equivalent to performance gains of 104 percent and 103 percent, respectively. According to Artificial Analysis’ Intelligence Index, the two rival models rank sixth and seventh globally among open-weight AI systems, placing LG’s model near the top tier.
LG said the model secured both performance and cost efficiency by applying proprietary technologies that improve inference efficiency over Exaone 4.0, released in mid-July, while reducing memory requirements and computational load.
K-Exaone boosts inference speed through architectural optimization, adopting a mixture-of-experts structure and hybrid-attention technology that cuts memory usage and computational demand by about 70 percent. Notably, the model runs on older GPU environments such as Nvidia’s A100, lowering deployment and operational costs and enabling easier adoption by startups and small and medium-sized enterprises.
LG AI Research said it completed development of K-Exaone in five months and plans to develop models with trillions of parameters going forward.
“K-Exaone has achieved its goal of outpacing the latest AI models from tech giants by more than 100 percent,” an LG official said. “We will continue to advance our proprietary technologies to strengthen performance and contribute to national competitiveness.”
andreahaneu@heraldcorp.com
