LG's Exaone Deep AI model schools DeepSeek's R1 in scholastic tests

Home > Business > Tech

LG's Exaone Deep AI model schools DeepSeek's R1 in scholastic tests

Published: 18 Mar. 2025, 16:34 Updated: 18 Mar. 2025, 16:55

JIN EUN-SOO
[email protected]

Audio report: written by reporters, read by AI

LG AI Research President Bae Kyung-hoon [LG CORP.]

LG AI Research has launched Exaone Deep, Korea's first inference AI model claimed to be able to reason and solve complicated problems in math, science and more.

The research arm of LG Corp. said Tuesday that it unveiled the open-source model at the GTC event organized by Nvidia in San Jose, California.

Exaone Deep 32B — a model with 32 billion parameters — outperformed DeepSeek's R1 in tests like the American Invitational Mathematics Examination (AIME) 2024 and AIME 2025 as well as Korea's, college entrance exam, the CSAT 2025, despite having one-fifth of the parameters of R1, which has 671 billion, according to a performance test result disclosed by open source AI platform Hugging Face.

On the Korean test, often referred to as suneung, Exaone Deep scored 94.5 to R1's 89.9.

In a subset of the Graduate-Level Google-Proof Q&A Benchmark, however, which is specifically designed to evaluate advanced reasoning of AI models in physics, biology and chemistry, the Exaone Deep model underperformed R1.

LG AI Research also noted that Exaone Deep was recognized by U.S.-based nonprofit AI research agency Epoch, making its list of Notable AI Models.

"We expect Exaone Deep to demonstrate high performance in specialized evaluation metrics for mathematics, science and coding, making it highly applicable not only in professional fields required by businesses but also in scientific research and educational settings including physics and chemistry," the company said in a release Tuesday.

The research arm also unveiled a lightweight model, Exaone Deep-7.8B, for on-device applications as well as an open-source 2.4B version.

Despite being 24 percent of the size of the 32B version, the light model was able to keep up with 95 percent of the larger one's performance, according to the company.

It added that the 7.8B version's performance surpassed that of OpenAI's equivalent model, o1-mini.

LG AI Research, along with Naver, is one of the pioneering firms in Korea seriously committed to the development of AI models.

After launching the first version of Exaone in 2021 as a generative AI model, the AI research arm of LG Corp. has been making constant updates to it in various ways.

In 2023, it launched the second version targeting "the top one percent of professionals" with an elite-level performance in areas like bio research and development, thesis research and image generation.

BY JIN EUN-SOO [[email protected]]