LG AI Research announced on the 18th that it will release the 'EXAONE Deep' to compete with global reasoning artificial intelligence models as open source. This release is significant as it demonstrates that a reasoning AI developed by a Korean corporation has proven its competitiveness in the global market. LG AI Research will introduce EXAONE Deep at NVIDIA's annual developer conference (GTC), held in San Jose, California, from the 17th to the 21st (local time).
LG AI Research unveiled its self-developed reasoning AI, 'EXAONE Deep,' signaling a transition to the 'Agentic AI' era. Agentic AI refers to proactive artificial intelligence that can autonomously make decisions through a process of setting hypotheses and conducting reasoning to verify them. Developing 'reasoning AI' that surpasses existing 'knowledge AI' is essential for this implementation.
Currently, only a few corporations worldwide, including the United States' OpenAI and Google, and China's DeepSeek and Alibaba, possess foundation models and have developed their own reasoning AI. 'EXAONE Deep' is the first domestic model that can compete with these entities.
LG stated that 'EXAONE Deep-32B' (with 32 billion parameters) demonstrated superior performance in comparisons with models from the United States and China, even with only 5% of the parameters of DeepSeek R1 (with 671 billion parameters). LG AI Research released various performance evaluation results assessing the logical reasoning and problem-solving abilities of three different EXAONE Deep models on the global open-source AI platform Hugging Face. It particularly proved excellence in solving complex mathematical and scientific problems.
EXAONE Deep-32B is based on the EXAONE foundation model, which has strengths in the Korean language, achieving a top score of 94.5 in the math section of the 2025 academic year college entrance exam and attaining a grade of 1 in all elective subjects (probability and statistics, calculus, geometry). It also scored 95.7 in the Math (MATH)-500, a measure of mathematical problem-solving ability.
It also achieved a score of 66.1 in the GPQA Diamond Test, which evaluates scientific problem-solving ability, marking a leading result in competition with reasoning AI models of similar parameter scale. The LiveCodeBench test for evaluating coding ability also yielded a score of 59.5, demonstrating competitive performance.
'EXAONE Deep-32B' has been recognized for its technological competitiveness by being listed among the notable AI models selected by the U.S.-based nonprofit AI research organization Epoch AI, alongside its open-source release. LG AI Research noted, 'EXAONE Deep' demonstrates high performance in specialized evaluation indicators in areas such as mathematics, science, and coding, and it is expected to be increasingly utilized not only in specialized fields required by corporations but also in scientific research and education settings in physics and chemistry.