The First Domestic Inference AI is Born, LG's 'EXAONE Deep' Unveiled… "Top Score in College Scholastic Ability Test Math"
LG AI Research Institute has made its "EXAONE Deep," a global inference AI model, open-source, positioning it as the first inference AI developed by a South Korean company to compete in the global market. From the 17th to the 21st (local time), LG AI Research is participating in NVIDIA’s annual developer conference (GTC) in San Jose, California, to introduce its self-developed inference AI, EXAONE Deep. LG AI Research is signaling a shift to the "Agentic AI" era through EXAONE Deep. Agentic AI refers to active AI that can independently generate and verify hypotheses through inference. Achieving this requires "inference AI," which goes beyond traditional "knowledge AI." Currently, only a few companies with foundation models, such as OpenAI and Google in the U.S. and DeepMind and Alibaba in China, have developed their own inference AI. EXAONE Deep is the first South Korean model capable of competing with these global players. ◆ Proven Excellence in Inference Problem Solving… 'Top Score in College Scholastic Ability Test Math' LG AI Research has released various performance evaluation results assessing the logical reasoning and problem-solving abilities of "EXAONE Deep." The "EXAONE Deep-32B" demonstrated excellence in solving complex math and science problems. Based on the EXAONE Foundation model, which is strong in Korean, EXAONE Deep-32B scored 94.5 in the 2025 College Scholastic Ability Test math section, achieving the highest score. It attained top marks in all elective subjects (Probability and Statistics, Calculus, and Geometry). In the math problem-solving metric, MATH-500, it scored 95.7, outperforming Alibaba’s QwQ32B (95.5) and the compressed distillation model DeepMind-R1-Llama-70B (94.3). R1 had the highest score at 97.3. It also achieved impressive results in evaluating scientific reasoning abilities. In the GPQA Diamond Test, which assesses scientific problem-solving skills, EXAONE Deep-32B scored 66.1, and in LiveCodeBench, which evaluates coding ability, it recorded a score of 59.5, showing competitive performance. LG AI Research anticipates that EXAONE Deep-32B will not only be useful in specialized fields required by companies but also gain significant applications in scientific research and education, such as physics and chemistry. The model has gained recognition for its technological competitiveness, being listed in the noteworthy AI models by Epoch AI, a non-profit AI research organization in the U.S., following its open-source release. ◆ Lightweight Model Released "Expansion in Smartphone and Robotics Industries" LG AI Research has also made the lightweight model "EXAONE Deep-7.8B" and the on-device model "EXAONE Deep-2.4B," developed alongside the "EXAONE Deep-32B," available as open-source. Despite being only 24% the size of the 32B, the lightweight model maintains 95% of its performance, and the on-device model, at just 7.5% of the original size, retains 86% of its performance, offering both cost-effectiveness and practicality. In fact, the performance of the lightweight model 7.8B surpasses that of OpenAI's o1-mini, and the on-device model 2.4B also demonstrated superior performance in comparisons with similar models. Notably, the on-device model can process data securely within the device without external server connections, making it advantageous in terms of security and privacy protection. The market anticipates rapid adoption of on-device models across various industries, including smartphones, automobiles, and robotics. An LG AI Research spokesperson commented, "After attending the Domestic AI Industry Competitiveness Review and Consultation meeting by the National AI Commission in February, where we announced that we would soon release a model equivalent to DeepSeek R1 as open-source, we unveiled EXAONE Deep just a month later. The core of LG's AI technology is to significantly reduce the model size while maintaining performance." ChatGPT를 사용하여 번역한 기사입니다.
