LG Unveils Korea's First Inference AI 'Exaone Deep'... "94.5 Points in CSAT Math"

First Unveiled at NVIDIA GTC
Top Performance with Just 5% the Size of DeepSeek
Record-High Score on 2025 CSAT Math
Outstanding Results in Coding, Physics, and Other Sciences
Lightweight and On-Device Models Also Released

LG AI Research Institute is set to release its inference-type artificial intelligence (AI) model, 'Exaone Deep,' as open source. Exaone Deep is the first domestic model capable of competing with global inference AI models such as OpenAI's 'ChatGPT.' Inference AI is an active AI that autonomously makes decisions by formulating hypotheses and conducting reasoning to verify them, signaling a transition to the era of 'Agentic AI.'

LG AI Research Institute announced on the 18th that it will unveil this model to the world at NVIDIA's annual developer conference (GTC), held from the 17th to the 21st (local time) in San Jose, California, USA.

Exaone Deep is an inference AI model introduced by LG to actively compete with ChatGPT, China's DeepSeek-R1, and others.

According to LG AI Research Institute, 'Exaone Deep-32B' (with 3.2 billion parameters) has only about 5% of the parameter scale of DeepSeek-R1 (671 billion parameters) but can deliver better performance than AI models from the US or China. The institute has released three Exaone Deep models on the global open-source AI platform 'Hugging Face.' Among them, 'Exaone Deep' uses a base model particularly strong in Korean and achieved a top score of 94.5 on the 2025 College Scholastic Ability Test (CSAT) math exam. It also received the highest grade in all elective subjects: Probability and Statistics, Calculus, and Geometry. The institute stated that this model demonstrated outstanding ability in solving math problems.

Exaone Deep-32B also showed excellent performance in the field of science. It scored 66.1 points on the 'GPQA Diamond Test,' which evaluates how well it solves doctoral-level physics, chemistry, and biology problems, outperforming other AI models of similar size. In addition to scientific problem-solving skills, it was recognized for its ability to utilize specialized knowledge. It also excelled in solving various problems such as coding, showing that it can effectively handle diverse scientific problems without complex coding.

LG AI Research Institute expects that Exaone Deep's high performance in specialized fields such as math, science, and coding will increase its utility not only in professional areas required by companies but also in scientific research and educational settings in physics and chemistry. Along with the open-source release, Exaone Deep-32B was listed as a 'Notable AI Model' by Epoch AI, a nonprofit AI research organization in the US. LG AI Research Institute also released the lightweight model 'Exaone Deep-7.8B' and the on-device model 'Exaone Deep-2.4B' as open source.

Baek Gyeonghun, Head of LG AI Research Institute. Photo by LG

The NVIDIA developer event 'GTC 2025,' attended by LG and others, was bustling as it opened. The San Jose Convention Center, where the event is held, saw lines of participants eager to experience cutting-edge technologies such as AI, robotics, and semiconductors from early morning. Attendees shared the latest technology information throughout the venue. In particular, the afternoon session on 'High Bandwidth Memory (HBM)' presented by Park Jung-soo, a researcher at SK Hynix, was packed with participants hoping to attend. About 30 participants were lined up at the session room door 40 minutes before the session started.

In the session, Researcher Park said, "We believe SK Hynix will lead technological and market innovation in HBM, but we are facing new technical challenges," adding, "We need to think of more efficient ways to improve bandwidth." He also expressed expectations for up to a 1.25% performance improvement through the 7th generation HBM4E, scheduled for mass production starting in the second half of next year. Furthermore, he emphasized the need for a new power distribution network (PBN) design and predicted that 'customized solutions' will become an industry trend in the future.

There is also great anticipation for CEO Jensen Huang's keynote speech. An event official emphasized, "The core of this GTC is how AI technology can be applied to robotics, especially physical aid for walking."