EXAONE
LG AI Research's family, published as open-weight checkpoints.
At a glance
Current release
From the release row K-EXAONE 2.0.
Family record: Official documentation · 5 Sep 2026
Built on it
No product in this catalogue records a dependency on this family yet. That is this directory’s coverage, not a statement about what is built on it.
Releases
-
EXAONE 3.0
First open EXAONE: a 7.8B instruction-tuned bilingual (English/Korean) model released on Hugging Face for research use.
-
EXAONE 3.5
Open-sourced in 2.4B on-device, 7.8B general and 32B sizes, announced with the ChatEXAONE enterprise agent.
-
EXAONE Deep
Reasoning models in 2.4B, 7.8B and 32B sizes for math, science and coding.
-
EXAONE 4.0
Hybrid model with non-reasoning and reasoning modes in 32B and 1.2B sizes, 131,072-token context, tool use, under a non-commercial licence.
-
EXAONE 4.5
First open-weight EXAONE vision-language model, 33B parameters (32B language model plus 1.2B vision encoder), 256K context, non-commercial licence.
-
K-EXAONE 2.0
750B-parameter mixture-of-experts model (A37B active) from Korea's sovereign foundation model project, moved to Apache-2.0, supporting 10 languages.