On March 18, 2025, LG AI Research unveiled Exaone Deep, South Korea’s inaugural reasoning artificial intelligence (AI) model. This development signifies a pivotal moment in the nation’s technological advancement, aligning it with global leaders in AI innovation. Unlike traditional AI systems that primarily rely on pre-learned data to generate responses, reasoning AI models, such as Exaone Deep, emulate human-like logical, step-by-step thinking processes to derive answers. This approach enhances the AI’s ability to tackle complex problems, particularly in fields requiring advanced reasoning capabilities.
Exaone Deep: An In-Depth Overview
The flagship model, Exaone Deep-32B, is equipped with 32 billion parameters. In the realm of AI, parameters are variables that the model learns during training, enabling it to make predictions or decisions based on input data. Generally, a higher number of parameters can enhance a model’s performance by allowing it to capture more intricate patterns in data. However, this also increases the computational resources required for both training and inference. Notably, DeepSeek’s R1 model contains 671 billion parameters, yet Exaone Deep-32B, with approximately 5% of that number, demonstrates comparable performance.
Performance Benchmarks and Comparative Analysis
In evaluations against leading reasoning AI models like DeepSeek-R1 and Alibaba’s QwQ-32B, Exaone Deep-32B has demonstrated exceptional capabilities, particularly in mathematical reasoning. In the 2024 U.S. Mathematical Olympiad, Exaone Deep-32B achieved a score of 90, surpassing DeepSeek-R1’s 86.7 and Alibaba’s QwQ-32B’s 86.7. Similarly, in South Korea’s 2025 College Scholastic Ability Test (CSAT) math section, it scored 94.5, the highest among its peers. Furthermore, it performed well on doctoral-level science problems, scoring 66.1, ahead of Alibaba’s QwQ-32B, which scored 63.3.
Despite its strengths in mathematical and scientific reasoning, Exaone Deep-32B has room for improvement in coding and language skills. In the Massive Multitask Language Understanding (MMLU) test, it scored 83, trailing behind Alibaba’s model, which scored 87.4, and DeepSeek’s model, which scored 90.8. Industry experts note that reasoning AI models are primarily designed for solving complex problems in math and science, which may result in less robust language capabilities compared to larger models.
Additional Models and Open-Source Initiatives
LG AI Research has also introduced the lightweight Exaone Deep-7.8B and the on-device Exaone Deep-2.4B models. The lightweight version retains 95% of the performance of the 32B model despite being just 24% of its size, while the on-device model achieves 86% performance at only 7.5% of the scale. To encourage developer engagement and foster innovation, LG has released the source code for these models as open-source, similar to DeepSeek’s approach.
Strategic Focus on Internal Utilization and Operational Costs
While the source code is freely available, LG intends to utilize the AI model internally for product development. Deploying it for public use, akin to OpenAI’s ChatGPT, would require massive data centers and entail substantial costs. This strategic decision allows LG to integrate the model’s capabilities into its products and services without incurring the significant expenses associated with public deployment.
South Korea’s AI Landscape and Collaborative Efforts
Other South Korean companies are also advancing in AI development. Naver, for instance, launched HyperCLOVA X in 2023 and has since reduced the number of parameters by about 60% while enhancing reasoning performance, leading to improved operating costs. The company is also developing a dedicated reasoning AI model. Additionally, AI startup Upstage has recently initiated full-scale development of reasoning AI, signaling a robust and competitive AI ecosystem within South Korea.
In a notable development, OpenAI has announced a partnership with South Korean chat app operator Kakao to jointly develop AI products. This collaboration aims to enhance Kakao’s products, including its dominant messaging app KakaoTalk, by integrating OpenAI’s advanced AI technologies. This partnership underscores South Korea’s commitment to integrating cutting-edge AI solutions across various sectors.
Global Context and Future Prospects
The introduction of Exaone Deep comes at a time when global tech giants are making significant strides in AI development. Companies like Baidu have launched models such as Ernie X1 and Ernie 4.5, positioning them as competitively priced alternatives to current market leaders. Ernie X1, for instance, rivals DeepSeek’s R1 at half the cost, while the multimodal Ernie 4.5 surpasses GPT-4.5’s performance for just 1% of the price. These developments reflect China’s endeavor to lead in the global AI arena by 2030.
LG’s introduction of Exaone Deep signifies a significant step in South Korea’s AI development, aligning the nation with global advancements in reasoning AI technology. As AI continues to evolve, the focus on creating models that can reason and think in a human-like manner will likely become more pronounced, leading to more sophisticated and capable AI systems in the future.
