K-EXAONE 2.0

model
LG

Korea's largest open-weights model: a 750B-total / 37B-active MoE upcycled from K-EXAONE 236B, covering 10 languages under Apache 2.0 — Phase 2 of the government's Sovereign AI project. Claims a 70.1 average over 24 benchmarks and 94.4 on OpenAI-MRCR; ships FP8, NVFP4, and "DSpark" speculative-decoding variants. Tech report is a PDF in the HF repo (no arXiv yet).

Announcement (LG AI Research)HuggingFaceGitHub

Model Details

Architecture MOE
Parameters 750B
Active params 37B
License Apache 2.0
Base model k-exaone
frontiermoeopen-weightmultilingual

Related