LG AI Research has officially launched Exaone 4.5, an advanced artificial intelligence model designed to expertly process both text and images. This significant release marks a crucial step in LG’s strategy to integrate robust visual capabilities into its flagship entry for South Korea’s highly competitive, government-backed sovereign AI initiative.
Recent internal benchmark tests conducted by LG AI Research demonstrate Exaone 4.5’s impressive performance. The multimodal AI model achieved an average score of 77.3 across five critical visual reasoning metrics within the demanding fields of science, technology, engineering, and mathematics (STEM). This places Exaone 4.5 ahead of OpenAI’s GPT-5 mini, which scored 73.5, and Anthropic’s Claude Sonnet 4.5, with a score of 74.6. It is important to note, however, that the benchmarked competitors represent mid-tier models within their respective company portfolios.
Contextually, OpenAI’s GPT-5 mini is characterized as a lightweight variant within their product suite, having been superseded by the more recent GPT-5.4 mini release in March. Furthermore, LG AI Research transparently notes that certain figures presented in its published benchmark tables are derived from self-measurement, rather than being subjected to independent third-party verification.
Despite its scale of 33 billion parameters, which is approximately one-seventh the size of K-Exaone – LG’s formidable 236-billion-parameter model actively contending in the Korean government’s prestigious Proprietary AI Foundation Model project – Exaone 4.5 achieves remarkable efficiency. LG highlights that, notwithstanding this substantial size disparity, Exaone 4.5 impressively matches K-Exaone’s text reasoning performance. This exceptional outcome is primarily attributed to its innovative hybrid processing architecture and advanced, faster inference techniques.
The overarching Proprietary AI Foundation Model project, meticulously overseen by South Korea’s Ministry of Science and ICT, represents a significant multi-year public-private partnership. This ambitious initiative is backed by approximately 530 billion won (equivalent to $358 million USD) and aims to cultivate globally competitive AI models by 2027. The rigorous selection process involves phased eliminations, with the two top-performing teams ultimately granted concentrated government GPU resources. K-Exaone distinguished itself by scoring highest across all categories in the project’s initial evaluation round held in January, and it is poised for a second review following August. LG positions Exaone 4.5 as a strategic preparatory step, crucial for the seamless integration of advanced vision capabilities into the larger K-Exaone model for this national competition.
In a move to foster broader AI innovation, Exaone 4.5 has been made available with open weights on Hugging Face, specifically intended for academic research and educational applications. However, it is important to note that commercial licensing for Exaone 4.5 currently remains restricted.
“With the foundational release of Exaone 4.5, our objective is to significantly expand AI’s perceptual and cognitive capabilities to encompass voice, video, and complex physical environments,” stated Lee Jin-sik, the distinguished head of the Exaone Lab at LG AI Research. “This progression is vital as we endeavor to develop sophisticated AI solutions capable of making robust, real-world decisions within diverse industrial settings.”
mjh
