Korea's Largest Model Goes Apache 2.0

LG AI Research released K-EXAONE 2.0 on Hugging Face on 31 July 2026. At 750 billion parameters it is the largest AI foundation model developed in Korea, more than three times the scale of the 236B first-phase model.

The repository ID - LGAI-EXAONE/K-EXAONE-2.0-750B-A37B - encodes the architecture: 750B total parameters with 37B active per token, a mixture-of-experts ratio close to GLM-5.2's 744B/40B. LG publishes BF16 alongside FP8 and NVFP4 quantised variants, which is what makes a 750B model deployable on hardware most organisations can actually get.

The significant part is the licence. LG moved the model to Apache 2.0, which permits commercial use without restriction. That is a meaningful change from the research-and-education-only terms attached to earlier EXAONE releases, and it puts K-EXAONE 2.0 in the small group of genuinely self-hostable frontier-scale models alongside Mistral Large 3, GLM-5.2, Nemotron 3 Ultra, Falcon H1 and Cohere's Command A+.

For context on LG's range, the company also released EXAONE 4.5 in April 2026: a 33-billion-parameter vision-language model roughly one-seventh the size of K-EXAONE, pairing a proprietary vision encoder with an LLM in a unified architecture. LG reported it achieving comparable text comprehension and reasoning performance to the much larger model, though that release went out under research, academic and educational terms rather than Apache 2.0.

K-EXAONE 2.0 is on our roster with its specification table already populated.