LG AI Research Institute unveiled its independently developed reasoning AI model ‘EXAONE Deep’ as open source on the 18th. This is assessed as a case where a reasoning AI developed by a domestic company has proven its competitiveness in the global market.
Through this release, LG AI Research aims to expedite the ‘Agentic AI’ era, where AI autonomously formulates and verifies hypotheses. In a situation where securing reasoning AI technology is essential to implementing AI capable of autonomous decision-making beyond traditional knowledge-based AI, ‘EXAONE Deep’ hints at full-scale competition with leading global companies.
LG AI Research showcased ‘EXAONE Deep’ at NVIDIA’s annual developer conference (GTC) held in San Jose, California, as a means of establishing its presence in the global AI market.
The ‘EXAONE Deep-32B’ model recorded high performance with only 5% the scale of the 671 billion parameter DeepSeek R1, based on 32 billion parameters. It demonstrated its ability to solve complex math and science problems, achieving a top grade with 94.5 points in the 2025 college scholastic ability test (CSAT) math section. In the MATH-500 index, which evaluates math problem-solving capabilities, it scored 95.7 points.
The scientific reasoning capability of the reasoning AI is also at a globally top level. It received 66.1 points in the GPQA Diamond Test, which assesses doctoral-level science problem-solving ability, marking the highest performance among AI models of its scale. In LiveCodeBench, an indicator of coding ability, it scored 59.5 points, indicating high applicability even in professional fields.
LG AI Research expects that EXAONE Deep will be widely utilized in industrial settings and educational and research fields relating to math, science, and coding. ‘EXAONE Deep-32B’ was also listed among the ‘Notable AI Models’ by Epoch AI, a non-profit AI research organization in the U.S., recognizing its technical capabilities.
In addition to ‘EXAONE Deep-32B’, LG AI Research also introduced the lightweight model ‘EXAONE Deep-7.8B’ and the on-device model ‘EXAONE Deep-2.4B’. The lightweight model reduced the size to 24% of the existing model while maintaining 95% of the performance, and the on-device model secured 86% of the performance with 7.5% of the size.
Especially, the on-device model allows data processing internally within the device, without an external server, providing strengths in terms of security and privacy protection. The introduction of on-device AI is expected to accelerate across various industries, including smartphones, cars, and robots.
An official from LG AI Research said, “The core of LG AI technology is reducing model size significantly while maintaining performance,” and added, “This announcement will be a turning point that changes the paradigm of the AI industry.”
LG Group plans to expand the AI ecosystem by linking AI technology advancements with various industries such as electronics, telecommunications, and automobiles. LG CEO Kwang-Mo Koo also emphasized AI-centric innovation in his New Year’s address, stating, “We will make AI conveniently usable in everyday life.”