LG EXAONE 4.0

EXAONE 4.0, unveiled July 2025 by LG AI Research, is Korea's first open-weight hybrid AI model: it fuses a chat-oriented LLM with a reasoning model so one checkpoint handles both rapid responses and hypothesis-driven problem solving. The 32B expert model posted 92.3/81.8 on MMLU-Redux/MMLU-Pro, 75.4 on GPQA-Diamond, and 85.3 on AIME 2025 in LG's evaluations, plus a 1.2B on-device sibling aimed at phones, cars, and robots. It ships with a 130K context window and passed six Korean professional licensing exams in LG's demos.

The honest picture: weights are open on Hugging Face but released for research and academic purposes — commercial licensing is murkier than Apache/MIT families. LG's benchmark framing is vendor-run and should be cross-checked on independent leaderboards. The Korean-language and professional-exam strengths are real but niche for non-Korean developers, and the ecosystem (fine-tunes, tooling) is thinner than Qwen or Llama.