LG AI Research Institute (Director Back -Hoon) unveiled its inference artificial intelligence (AI) model ‘EXAONE Deep’ on the 18th.
The self -reasonable model stressed that only a small variety of firms with foundation models comparable to Open AI, Google, China Deep Chic and Alibaba are developing their very own inference AI. Exa1 Dip is the primary model in Korea to compete with them.
‘Exa1 Deep-32B’ explained that it showed excellent performance at 5% of ‘Deep Chic-R1’, which has 670 billion parameters.
In complex mathematics (AIME 2024, AIME 2025), R1 showed the identical performance. The 2025 SAT (CSAT 2025) scored 94.5 points within the mathematics area, and achieved the primary grade in each elective courses (probability and statistics, calculus, geometry).
The R1 was behind within the Live Code bench, which evaluates the scientific reasoning ability of Ph.D., comparable to physics, chemistry, and biology, and a live code bench that assesses coding skills.
LG AI researcher expects that Exa1 Dip will increase its utilization in various research and academic fields, as Exa1 Dip has high performance in specialized indicators comparable to mathematics, science and coding.
Specifically, Exa1 Deep-32B was listed on the list of EPOCH AI, an American non-profit AI research institute, concurrently the open source release.
Then again, on this present day, the small language model (SLM) ‘Exa One Deep -7.8B’ and ‘Exa1 Deep-2.4B’ for On-Vice were also released.
SLM emphasized that the performance is as much as 95% of the dimensions of 32B, which is about 95%, and the ondi -vice model is 7.5% and the performance is 86%. It’s explained that it has each high economics and practicality at the identical time.
An official of the LG AI researcher said, “We participated in a gathering held on the National AI Committee last month and announced that it should open a deep-chic-R1 model as an open source.
The truth is, the performance of SLM 7.8B exceeded the performance of ‘O1-Mini’ of Open AI, and the on-die vise model 2.4B also showed the world’s best performance amongst its classes.
As well as, the on -vise model is predicted to be utilized in various industries comparable to mobile phones, automobiles and robots due to its strength when it comes to security and private information protection.
LG AI researchers said they may lead the on -vise AI market by advancing models with affiliates comparable to LG Electronics and LG Uplus.
By Jang Se -min, reporter semim99@aitimes.com
