LG AI Research's Exaone is the company's proprietary artificial intelligence model series, first launched in December 2021. (LG AI Research)
LG AI Research's Exaone is the company's proprietary artificial intelligence model series, first launched in December 2021. (LG AI Research)

LG AI Research on Thursday released Exaone 4.5, an artificial intelligence model that can process both text and images, as the company builds toward adding visual capabilities to its entry in South Korea's government-backed sovereign AI competition.

In internal benchmark tests, LG said Exaone 4.5 scored an average of 77.3 across five visual reasoning metrics in science, technology, engineering and math, ahead of OpenAI's GPT-5 mini at 73.5 and Anthropic's Claude Sonnet 4.5 at 74.6. The comparison targets, however, are mid-tier models from their respective companies.

Exaone 4.5 averaged 77.3 across five STEM visual reasoning benchmarks, ahead of GPT-5 mini at 73.5 and Claude Sonnet 4.5 at 74.6, according to LG AI Research's internal tests. (LG AI Research)
Exaone 4.5 averaged 77.3 across five STEM visual reasoning benchmarks, ahead of GPT-5 mini at 73.5 and Claude Sonnet 4.5 at 74.6, according to LG AI Research's internal tests. (LG AI Research)

GPT-5 mini is a lightweight variant in OpenAI's lineup, having already been succeeded by GPT-5.4 mini in March. Some figures in LG's published benchmark tables are marked as self-measured rather than independently verified.

At 33 billion parameters, the model is roughly one-seventh the size of K-Exaone, LG's 236-billion-parameter model competing in the Korean government's Proprietary AI Foundation Model project. Despite the size gap, LG said Exaone 4.5 matches K-Exaone's text reasoning performance, a result it attributed to a hybrid processing architecture and faster inference techniques.

The Proprietary AI Foundation Model project, overseen by the Ministry of Science and ICT, is a multiyear public-private effort backed by roughly 530 billion won ($358 million) to develop globally competitive AI models by 2027. Teams are eliminated in phases, with the final two receiving concentrated government GPU resources. K-Exaone scored highest across all categories in the project's first evaluation round in January and faces a second review after August. LG described Exaone 4.5 as a preparatory step toward integrating vision capabilities into that larger model.

Exaone 4.5 has been released as open weights on Hugging Face for research and educational use. Commercial licensing remains restricted.

"Starting with this model, we will expand AI's understanding to voice, video and physical environments to build AI that makes real decisions in industrial settings," said Lee Jin-sik, head of the Exaone Lab at LG AI Research.


mjh@heraldcorp.com