K-EXAONE 2.0 is a frontier-scale multilingual open-weight language model from LG AI Research. It uses an MoE design that activates 37B of 750B total parameters per token — 78 layers (2 dense + 76 sparse) routing 8 of 256 experts — combined with hybrid attention that mixes a global NoPE layer with 4,096-token sliding-window attention. It offers a 262,144-token context, a built-in reasoning (extended thinking) mode, and structured tool calling optimized for agentic workflows, plus two speculative decoding methods (MTP and DSpark) delivering roughly 3-5x faster generation. It shows particular strength in long-context retrieval (OpenAI-MRCR 94.4%), safety (KGC-Safety 99.8%), and agentic coding (SWE-Bench 68.2%), and supports ten languages including Korean, English, and Japanese. Upcycled from the original K-EXAONE via continual pretraining and difficulty-focused mid-training, it is released under Apache 2.0 for commercial use.
심층 사고|공개 모델Apache-2.0
전문가 성능 평가
추론·수학
⚠
GPQA Diamond
82%↓3%
⚠
HLE
18%↓10%
⚠
MMLU-Pro
84%↑1%
코딩
⚠
SciCode
37%↓6%
언어·지시
⚠
IFBench
73%↑8%
⚠
AA-LCR
56%↓13%
K-EXAONE