LG AI Research has introduced EXAONE Deep, an open-source reasoning AI model. The model was unveiled on March 18 and showcased at NVIDIA’s GTC conference in San Jose, California.
LG AI Research asserts that EXAONE Deep outperforms top AI models in scientific comprehension and mathematical reasoning, including Google DeepMind’s Gemini, China’s DeepSeek, and OpenAI’s GPT series. Engineered for advanced problem-solving in mathematics, science, and coding, the model enters a field long dominated by a handful of major players.
The Agentic AI independently sets hypotheses and verifies them through reasoning. Unlike conventional knowledge-based AI, Agentic AI enhances autonomy in decision-making. With this milestone, EXAONE Deep becomes Korea’s first reasoning AI built on a foundation model.
The flagship EXAONE Deep-32B model, with 32 billion parameters, achieves world-class performance while operating at just 5% of the size of competing models like DeepSeek R1 (671B parameters).
The model demonstrated problem-solving skills, scoring 94.5 in the 2025 CSAT mathematics test and achieving the highest grades in elective subjects like calculus and geometry. It also excelled in international benchmarks, including a 90.0 score on AIME 2024 and a 95.7 score in MATH-500, proving its ability to handle complex mathematical challenges.
Beyond mathematics, EXAONE Deep showcases strong scientific reasoning and coding capabilities. It recorded 66.1 points on the GPQA Diamond test, which assesses doctoral-level problem-solving in physics, chemistry, and biology, and achieved 59.5 points on LiveCodeBench, an index evaluating coding proficiency.
LG has also released lighter versions of EXAONE Deep to balance performance with efficiency. The EXAONE Deep-7.8B model retains 95% of the 32B model’s performance while being only 24% of its size, while the on-device 2.4B model delivers 86% of the performance at 7.5% of the size. The on-device model is particularly beneficial for secure, private AI applications in smartphones, automobiles, and robotics, as it processes data internally without connecting to external servers.
Recognizing its technological advancements, Epoch AI, a U.S. non-profit research organization, listed EXAONE Deep among its ‘Notable AI Models.’ LG is the only Korean entity to receive this recognition two years in a row.
LG AI Research aims to dominate the on-device AI sector by refining its models in collaboration with LG affiliates like LG Electronics and LG U+. Chairman and CEO Koo Kwang-mo highlighted LG’s commitment to innovation, stating that the company has continually pushed boundaries to create new value. He emphasized LG’s vision of integrating advanced AI into everyday life, enabling people to focus on more meaningful and enjoyable activities.
The EXAONE Deep models have demonstrated exceptional performance across multiple domains, particularly in physics, chemistry, biology, and coding. The 7.8B and 2.4B versions outperformed their global counterparts, including OpenAI’s o1-mini, in lightweight and on-device AI tasks. LG AI Research sees these advancements as a step toward AI systems that can handle complex challenges and enhance daily life through continuous innovation.
