LG AI Research Open-Sources K-EXAONE 2.0: Korea's Largest AI Model at 750B Parameters

Summary: LG AI Research released K-EXAONE 2.0 on July 31 — a 750B-total, 37B-active-parameter hybrid-attention Mixture-of-Experts model available on Hugging Face under the permissive Apache 2.0 license.

Key facts

• Architecture: hybrid-attention MoE with 750B total / 37B active parameters per token, enabling frontier-scale reasoning at a fraction of the inference cost of dense models.
• Supports 10 languages including Korean, English, Japanese, Chinese, Spanish, and German, with a 262,144-token context window.
• Released under Apache 2.0 — commercial use permitted, weights freely downloadable from Hugging Face.
• Average benchmark score improved from 63.3 to 70.1, a gain of more than 10% over its predecessor.

Why it matters

K-EXAONE 2.0 is South Korea's clearest demonstration yet that a domestic lab can independently build and open-source a frontier-class model. Arriving two days after another 700B-scale open release, it signals that non-US players are accelerating in the race for sovereign AI capability. The Apache 2.0 license makes it immediately usable for commercial products, broadening the base model options for Korean AI startups and enterprises.

Read more

• LG unveils K-EXAONE 2.0 — Korea Times
• Hugging Face model card — LGAI-EXAONE