LG Launches Korea’s First Reasoning AI Model to Compete with OpenAI and DeepSeek
Introduction to LG Exaone Deep: Korea’s First Reasoning AI Model
LG AI Research has recently launched Exaone Deep, the first reasoning artificial intelligence model in South Korea. This innovative model aims to compete with industry leaders like OpenAI and DeepSeek. The announcement was made during the Nvidia GTC conference in Silicon Valley, where the model was introduced as a groundbreaking achievement in AI technology.
Key Features of Exaone Deep
Performance and Efficiency
The Exaone Deep model boasts an impressive 32 billion parameters, a fraction of the 671 billion parameters found in China’s DeepSeek R1. Despite this, LG claims that Exaone Deep offers superior performance, particularly in tackling complex scientific and mathematical issues. The model is designed for efficiency, operating on just one unit of Nvidia’s H100 chip, compared to the 16 chips required by its competitors like DeepSeek. This efficiency allows for lower operational costs and higher learning capabilities.
Benchmark Scores
Exaone Deep’s performance has been validated through various academic and technical assessments. Listed below are its notable achievements:
- Korean College Scholastic Ability Test (CSAT) Mathematics: Scored 94.5, outperforming OpenAI’s o1-mini (84.4) and DeepSeek R1 (88.8).
- American Invitational Mathematics Examination 2025: Achieved a score of 80, rivaling DeepSeek R1.
- Graduate-level Google-Proof Q&A Benchmark: Scored 66.1, showcasing its problem-solving abilities in fields like physics, chemistry, and biology.
- LiveCodeBench: Registered a score of 59.5, excelling among similar-sized AI models.
These results highlight Exaone Deep’s strong capabilities in both scientific challenges and coding tasks.
Understanding Reasoning AI
The Importance of Inference Capability
According to LG AI Research, a key feature of reasoning AI is its ability to think and draw conclusions—similar to human thought processes. A research official noted that enhancing inference capability is essential for developing AI agents and robots that meet human needs effectively.
The Model’s Variants
In addition to the main Exaone Deep model, LG also introduced two other variations:
Lightweight Model (7.8B):
- Size: 24% of the 32B model
- Performance: Maintains 95% of the larger model’s capabilities.
- On-Device Model (2.4B):
- Size: Just 7.5% of the original model
- Performance: Achieves 86% of the 32B model’s performance.
These models complement the main offering, making advanced AI technology more accessible for on-device applications while still delivering robust performance.
Acclaim and Recognition
Exaone Deep was quickly recognized after its launch, being featured on Epoch AI’s Notable AI Models list, making it the only Korean model to receive such an honor in the past two years. This recognition underscores the impact of LG’s investment in research and development.
Open Source Initiative
In a move to foster innovation and collaboration, LG AI Research has made all versions of Exaone Deep available as open-source. This decision allows researchers, developers, and enthusiasts to access, modify, and enhance the models, promoting a collective advancement in AI technology.
Vision for the Future
Koo Kwang-mo, CEO of LG Corp., emphasized the company’s ongoing commitment to pioneering new technologies that enhance everyday life. He envisions a future where advanced technologies like AI become integrated into daily routines, providing convenience and freeing up time for more meaningful pursuits.
This launch marks a significant moment in the landscape of artificial intelligence, positioning Korea as a growing player in the global AI arena with cutting-edge reasoning technology.
