LG’s K-Exaone ranks in global top 10 AI models, offering efficient open-weight access

LG AI Research’s foundational model K-Exaone has entered the global top 10, ranking seventh on the Intelligence Index compiled by Artificial Analysis. K-Exaone is the only Korean model in the top 10 amid a list led by China (six models) and the US (three models). Developed over five years, the model uses a mixture-of-experts (MoE) architecture with 236 billion parameters (about 23 billion activated per inference) and hybrid attention to cut compute demands by roughly 70% versus previous models. LG reports K-Exaone topped 10 of 13 benchmark tests with an average score of 72 and scored 97.38 on LG’s KGC-Safety benchmark, outperforming OpenAI’s GPT-OSS-120B and Alibaba’s Qwen-3-235B on safety metrics. LG published K-Exaone as an open-weight model on Hugging Face and briefly reached second place on the platform’s global model trend chart. Free API access is available through January 28 to encourage developer uptake. Technical improvements include a 150,000-word tokenizer vocabulary, multi-token prediction boosting inference speed by 150%, and optimizations that make the model runnable on A100-class GPUs rather than top-tier hardware — lowering costs and broadening access. LG emphasised pre-training on reasoning trajectories to improve problem-solving and ran internal compliance and ethics reviews to remove potentially copyrighted material. Primary keywords: K-Exaone, LG AI Research, foundational model, mixture-of-experts, hybrid attention. Secondary/semantic keywords: Hugging Face, open-weight model, A100 GPUs, AI benchmarks, model safety. This development signals improved accessibility to frontier-class AI and strengthens Korea’s presence in global AI competition.