LG K-Exaone

K-EXAONE is LG AI Research's flagship foundation model — a 236B-parameter mixture-of-experts model with ~23B parameters active per inference, released as open weights on Hugging Face in January 2026. It's the product of five years of in-house research and Korea's entry into the frontier-model race: it ranked 7th globally on the Artificial Analysis Intelligence Index, the only non-US, non-Chinese model in the global top 10, and briefly hit #2 on Hugging Face's trending chart. Its hybrid attention architecture (sliding-window + global attention) cuts memory and compute requirements by 70% versus Exaone 4.0.

The honest picture: this is a first-phase model — LG's own lab head said it was trained on roughly half their available data, so expect stronger successors. As a Korean lab's model, English-language tooling, docs, and community support lag the US/Chinese ecosystems. And while the MoE design is efficient, 236B total parameters still demands serious inference hardware. It's a genuine achievement for Korean AI sovereignty, but pragmatically it's a research and evaluation play for most developers outside LG's ecosystem.