LG Artificial Intelligence (AI) Research head Bae Kyung-hoon speaks during LG AI Insight 2024 in western Seoul, Nov. 21, 2024. / Courtesy of LG Corp.
LG Artificial Intelligence (AI) Research has released its own reasoning AI model, Exaone Deep, as open source, the research lab announced Tuesday, signaling a heated competition with advanced AI models developed by OpenAI, Google and other global AI leaders.
According to the lab, it unveiled Exaone Deep at Nvidia's annual GPU Technology Conference (GTC) in San Jose, California, which started on Monday and will run through Friday.
In the context of AI, reasoning refers to a logical process of using knowledge to solve problems, draw conclusions and make predictions. Applying this reasoning to solve a specific problem is called inference.
Reasoning AI models are considered a critical prerequisite to realizing an agentic AI, which is capable of making autonomous decisions by independently formulating hypotheses and conducting inferences to verify them.
Currently, OpenAI and Google in the U.S. and DeepSeek and Alibaba in China are leading the market with foundation models to develop reasoning AI models, and LG's Exaone Deep is the first model developed in Korea that can compete with them.
In performance evaluation reports on the global open-source AI platform Hugging Face, Exaone Deep models demonstrated comparable performance in assessing logical reasoning skills and problem-solving capabilities, despite the Exaone Deep-32B model having only 32 billion parameters — just 5 percent of DeepSeek's R1, which has 671 billion parameters. Parameters are learned values that a model uses to analyze data and are widely regarded as an indicator of the model's scale or size.
A comparison between LG Exaone Deep-32B model and other reasoning AI models / Courtesy of LG Corp.
LG AI Research said Exaone Deep-32B proved its excellence in solving complex mathematical and scientific problems, particularly in the Korean language. The model recorded the highest score of 94.5 in this year's mathematics exam of Korea's College Scholastic Ability Test and achieved the top grade in all elective subjects. It also recorded 95.7 points in MATH-500, an index evaluating mathematical problem-solving capability.
The model also achieved top performance in the Google-Proof Q&A Benchmark diamond test, which evaluates graduate-level scientific reasoning capabilities in physics, chemistry and biology, recording 66.1 points and thereby outperforming reasoning AI models of a similar parameter scale. It also showed competitive performance with 59.5 points in LiveCodeBench, which evaluates coding ability.
LG AI Research said that it expects Exaone Deep will be widely used in professional and scientific fields. Exaone Deep-32B was listed on the Notable AI Models list selected by the U.S. non-profit AI research organization Epoch AI upon its open-source release.
Along with the 32B model, the lab revealed the lightweight Exaone Deep-7.8B and the on-device Exaone Deep-2.4B as open source. The lightweight model retains up to 95 percent of the 32B model's performance despite being only 24 percent of its size, and the on-device model achieves 86 percent of the performance at just 7.5 percent of its size.
“Just a month after announcing that we would release an AI model whose performance is on par with DeepSeek R1, we now proudly present Exaone Deep,” an LG AI Research official said. “The core of LG's AI technology is maintaining performance while significantly reducing model size.”
LG AI Research is LG Group's think tank for AI business, developing Exaone as the foundation model for multiple variation models such as LG Uplus' ixi-series and LG Electronics' devices. The lab has been expanding its presence in the global AI industry under LG Group Chairman Koo Kwang-mo's drive.
In his New Year address, Koo said, “We will create a new lifestyle where people use cutting-edge technologies such as AI conveniently in their daily lives, allowing them to spend their precious time on more enjoyable and meaningful tasks.”