Unified LLM integrating non-reasoning and reasoning modes via Hybrid Attention (sliding-window + global), reducing memory usage by 70%. The 32B variant achieved the highest Artificial Analysis Intelligence Index score of any 32B-class model.

Model Details

Architecture DENSE
Parameters 32B

Variants

Name Parameters Notes
EXAONE-4.0-1.2B 1.2B
EXAONE-4.0-32B 32B

Paper

arXiv: 2507.11407

open-weightreasoningefficiency

Related