LG AI Research said on Saturday it unveiled its proprietary AI foundation model, K-EXAONE. It said K-EXAONE consolidates foundation model technology it has accumulated over five years. It said it advanced hybrid attention technology to cut memory requirements and computation by 70 percent versus EXAONE 4.0.

It combines sliding window attention and global attention. It expanded its training vocabulary to 150,000 through tokenizer upgrades. It can handle documents 1.3 times longer than existing models. Inference speed improved by 150 percent versus existing models.

LG also said its K-EXAONE scored an overall average of 72 points across 13 benchmark tests in the first evaluation of a government proprietary AI foundation model project.

K-EXAONE scored an average of 97.83 points across four categories in KGC-SAFETY, a metric evaluating Korean-specific characteristics. That is higher than 92.48 points for U.S. OpenAI's GPT-OSS 120B model and 66.15 points for China's Alibaba Qwen3 235B model. LG AI Research will provide the K-EXAONE API for free until Jan. 28. It is conducting joint research with Seoul National University, KAIST and the University of Michigan, among others.

K-EXAONE can process a 260,000-token context at once, the longest among AI models in South Korea. That amounts to more than 400 A4 pages. It runs in an Nvidia A100-class GPU environment through an energy-efficient design. LG AI Research said it verified K-EXAONE's safety and reliability through its in-house AI ethics committee. It explained that it is conducting data compliance evaluations for all training data.

K-EXAONE scored 32 points in an intelligence index assessment by global AI performance evaluation organisation Artificial Analysis. It ranks seventh globally and first domestically among open-weight models. The current global top 10 open-weight models comprise six from China and three from the United States. It said K-EXAONE is the only model from South Korea to make the list.

K-EXAONE ranked second in global model trend rankings shortly after it was released on global open-source AI platform Hugging Face. It was also included in a list of notable AI models by U.S. nonprofit AI research institute Epoch AI. LG AI Research has placed five models on the list, the most among South Korean companies, from EXAONE 3.5 in 2024 through EXAONE Deep, EXAONE Path 2.0, EXAONE 4.0 and K-EXAONE.

Choi Jung-kyu, head of the Agentic AI Group at LG AI Research, said K-EXAONE shows that proprietary technical design can compete on equal terms with global large-scale models despite resource constraints. He said the institute will stay focused on research and development with confidence in developing South Korea's representative AI and build models that contribute to the development of the global AI ecosystem.