A graph evaluating 5 nationwide AI basis challenge fashions / Courtesy of LG AI Analysis
LG AI Analysis’s newly developed synthetic intelligence (AI) mannequin, Okay-EXAONE, created for the nationwide AI basis mannequin challenge, has demonstrated robust competitiveness in each home and worldwide AI arenas.
The corporate shared that Okay-EXAONE topped 10 out of 13 benchmark checks within the first-round analysis of the government-backed AI basis mannequin initiative, scheduled to conclude this month.
The mannequin earned a mean rating of 72 factors, delivering the strongest total efficiency amongst fashions developed by 5 consortia taking part within the challenge.
Okay-EXAONE scored 32 factors on the Synthetic Evaluation Intelligence Index, rating first in Korea and seventh worldwide amongst open-weight fashions, based on the worldwide AI evaluation platform.
The Korean mannequin was additionally the one non-Chinese language or U.S. entrant within the prime 10, which was dominated by six Chinese language and three U.S. fashions.
“We deliberate the mannequin’s improvement below sensible timelines and infrastructure constraints, utilizing solely about half of our obtainable coaching knowledge to construct Okay-EXAONE,” stated Lee Jin-sik, chief of LG AI Analysis’s EXAONE Lab.
“This preliminary launch is a place to begin towards a frontier-grade mannequin, and we plan to considerably scale up efficiency from this level onward.”
After releasing its open weights on the Hugging Face, Okay-EXAONE climbed to No. 2 on the platform’s world mannequin development rating, drawing robust curiosity from AI researchers worldwide.
The mannequin was additionally acknowledged by the U.S. nonprofit AI analysis institute Epoch AI for its Notable AI Fashions listing. With this, LG AI Analysis has turn into the primary Korean firm to put 5 fashions on the listing, together with EXAONE Deep and EXAONE 4.0.
The listing is extensively referenced in Stanford College’s annual AI experiences and is used as a benchmark for assessing nationwide and company AI competitiveness.
Okay-EXAONE represents the 5 years of LG AI Analysis’s in-house basis mannequin R&D, integrating its proprietary applied sciences to attain increased effectivity at decrease coaching and operational prices. Fairly than merely scaling by knowledge measurement, the crew re-engineered the mannequin structure itself.
The mannequin incorporates a complicated model of hybrid consideration that determines which data an AI mannequin ought to deal with when processing large datasets.
LG AI Analysis refined its hybrid consideration mechanism by combining sliding-window consideration for native focus with world consideration for full-context understanding. This strategy cuts reminiscence use and computational load by 70 p.c in comparison with the earlier mannequin, Exaone 4.0.
Okay-EXAONE has a complete parameter rely of 236 billion and adopts a mixture-of-experts structure, with solely about 23 billion parameters activated at a time. Parameters are the inner variables that an AI mannequin learns from knowledge throughout coaching to make predictions.
“Okay-Exaone’s environment friendly structure delivers frontier-level efficiency even on A100-class graphics processing items, relatively than relying on ultra-expensive infrastructure. This design will permit resource-limited organizations to deploy highly effective massive fashions, serving to increase the home AI ecosystem,” an LG official stated.
