LG AI Research has unveiled the country’s first reasoning artificial intelligence model, Exaone Deep with top-tier performance to compete head-on with industry-leading players including OpenAI and DeepSeek, Tuesday.
The research institute, affiliated with Korea's fourth-largest conglomerate LG, introduced the new foundation model at this year's Nvidia GTC, an AI conference held in Silicon Valley. The company claimed the model outperforms leading US and Chinese AI models with fewer parameters and graphic processing units.
According to LG, Exaone Deep 32B features 32 billion parameters, which is only 5 percent of the parameters of China's DeepSeek R1 -- runs on 671 billion parameters -- but demonstrates exceptional performance, especially in solving complex mathematical and scientific problems. A parameter is a numeric value that helps AI models process data and make decisions.
"Exaone Deep shows excellence in high-difficulty benchmarks, outperforming significantly larger models. In terms of the number of GPUs each model uses, Exaone Deep 32B runs on only one unit of Nvidia's H100 chip, while DeepSeek operates on 16 of those expensive chips," an LG AI research official explained, highlighting the learning efficiency and cost-effectiveness.
"Our reasoning AI model shows how Korea's AI technology is globally competitive, given how only a few players -- mainly in the US and China -- have developed in-house reasoning AI."
Exaone Deep 32B model demonstrated competitive superiority in science and coding domains, scoring higher than its rivals. The model achieved 94.5 points in the 2025 Korean College Scholastic Ability Test mathematics section when OpenAI o1-mini scored 84.4 points and DeepSeek R1, 88.8 points, in the data provided by the research center.
On the American Invitational Mathematics Examination 2025, Exaone Deep 32B achieved performance on par with the DeepSeek R1, with both models scoring 80 points.
The 32B model also achieved a 66.1 score on the Graduate-level Google-Proof Q&A Benchmark that evaluates PhD-level problem-solving abilities in physics, chemistry and biology. In LiveCodeBench, the benchmark for evaluating coding capabilities, it scored 59.5 points, outperforming other reasoning AI models of similar size, LG said.
"Humans think, reason and produce results. For AI agents and robots to evolve into what humans truly want and satisfy demands to the level we want, inference capability is the enabler," an LG official said.
Immediately after its release, the Exaone Deep 32B model was listed on Epoch AI’s Notable AI Models, marking another achievement following Exaone 3.5 and making the brand the only Korean model to be listed in the past two years, the company said.
In the unveiling, LG also introduced the lightweight model Deep 7.8B and on-device model 2.4B, which feature fewer parameters but prove their practical utility.
Being 24 percent the size of the 32B model, the lightweight model maintains 95 percent of its performance, while the on-device model achieved 86 percent of the big model's performance at just 7.5 percent the size, the company said.
Being Korea's first reasoning AI foundation model, Exaone Deep also delivers exceptional performance in general language understanding, securing the highest score of 83 in the Massive Multitask Language Understanding benchmark, an LG official said, noting how the lightweight 7.8B model outperformed the OpenAI o1-mini in the US.
"The core of LG's AI technology is reducing model size while maintaining performance.”
LG AI Research released all Exaone Deep models as open-source, allowing researchers, developers and other users to access and build on them.
Meanwhile, Koo Kwang-mo, CEO of LG Corp., emphasized in his New Year's address this year, "The current LG has been built by accumulating many moments of challenging new areas and creating unprecedented value."
"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," Koo said.
herim@heraldcorp.com
