LG AI Research unveils K-EXAONE 2.0 with 750 billion parameters

LG AI Research on Friday released K-EXAONE 2.0 on the open-source platform Hugging Face, ahead of the second evaluation round of the Ministry of Science and ICT's proprietary AI foundation model project.

K-EXAONE 2.0 is Korea's largest AI foundation model, with 750 billion parameters — more than three times the scale of the first-generation model, which had 236 billion. A higher parameter count generally enables a model to handle more complex tasks.

LG AI Research also switched the model's license to Apache 2.0, allowing anyone to use it commercially and broadening its accessibility.

The lab released benchmark results alongside the model, covering 24 evaluation metrics across nine categories — including those used in the government project's performance assessment. K-EXAONE 2.0 averaged 70.1 points across the 24 metrics, more than 10 percent above the first-generation model's score of 63.3. Coding and agentic coding stood out, with the average of three key metrics in those areas rising 30 percent.

The model matched or outperformed leading global models in key areas including long-context understanding, agentic tasks and instruction following. LG AI Research said it also invested significant effort in improving safety, ensuring compliance with Korea-specific standards and global ethical guidelines.

K-EXAONE 2.0 now supports 10 languages, adding French, Italian, Portuguese and Polish to the existing lineup of Korean, English, Spanish, German, Japanese and Vietnamese.

"This makes it the domestic AI foundation model supporting the most languages, laying the groundwork for sovereign AI to expand beyond Korea into global markets," an LG AI Research official said.

LG AI Research said it successfully completed large-scale training of a 750-billion-parameter model — a first in Korea at that scale — and has built out the full system infrastructure for model evaluation and inference. The key achievement, the lab said, is demonstrating real-world performance across diverse scenarios and marking the start of a serious bid to compete with global frontier models.

The lab is finishing development of a public evaluation site for real-world usability testing as part of the second review stage of the government's proprietary AI foundation model project, and is also preparing a service that will let the general public try K-EXAONE 2.0.

LG AI Research has been accelerating K-EXAONE's expansion across manufacturing, biotech, finance and the public sector. In April, it released EXAONE 4.5, a vision-language model selected as the AI engine behind the Ministry of Interior and Safety's AI Safety Report system, where it handles analysis of more than 39,000 safety reports per day.

The lab plans to release an additional industry-specific AI foundation model in early August, further building out its portfolio of specialized AI for industrial and everyday applications.

"K-EXAONE 2.0 is significant not simply because we built a model with a large parameter count, but because Korean researchers independently completed the entire process — from designing the 750-billion-parameter model to data training, distributed learning and inference environment setup," said Lim Woo-hyung, co-director of LG AI Research. "We have secured the capability to compete on equal footing with global frontier models, which is the core goal of the proprietary AI foundation model project."