Nota AI Successfully Optimizes 236B-Parameter Large AI Model on Korean NPU…“Maintains Original-Level Performance While Reducing Model Size by 71%”

▶ Successfully optimized LG AI Research’sK-EXAONE 236B on FuriosaAI’s data center NPU
▶ Reduced model size by approximately 71%, easing memory requirements for largeAI model deployment
▶ Maintained approximately 99.2% accuracy compared to the original model acrossthree major benchmarks
▶ Demonstrated the potential of sovereign AI infrastructure through thecombination of a Korean NPU, domestic AI model, and AI optimization technology
‍

Nota AI, an AI model compression andoptimization company led by CEO Myungsu Chae, announced that it hassuccessfully optimized LG AI Research’s flagship AI model, K-EXAONE 236B, onFuriosaAI’s data center neural processing unit (NPU).

K-EXAONE 236B is a large-scale AI modelwith approximately 236 billion parameters and is built on a Mixture-of-Experts(MoE) architecture, which selectively activates multiple expert models. WhileMoE structures can improve the efficiency of large models, they require highlysophisticated optimization to ensure that each expert model operates reliably.In particular, frontier-scale models often go through long reasoning processeswhen solving complex problems, meaning even small errors introduced duringquantization can accumulate and affect the accuracy of the final output. Thisachievement is significant in that Nota AI optimized such a large model for aKorean NPU environment while maintaining accuracy across major evaluations.

In this project, Nota AI optimized K-EXAONEfor FuriosaAI’s data center NPU. Rather than readjusting the entire model, NotaAI precisely analyzed specific sections where performance degradation couldoccur and applied optimization only where necessary to minimize performanceloss. As a result, the company enabled efficient execution of a large AI modelon a Korean NPU while maintaining performance at a level comparable to thebaseline model across key metrics.

The performance evaluation also showedmeaningful results. Nota AI reduced the size of K-EXAONE by approximately 71%,lowering the memory burden required to run a large AI model, while maintainingaccuracy close to the original model across key evaluation categories,including scientific reasoning, instruction following, and mathematical problemsolving. This demonstrates the potential to run a 236-billion-parameter modelmore efficiently and improve the operational efficiency of data center AIinfrastructure.

In Nota AI’s internal evaluationenvironment, the optimized model recorded 79.80 on GPQA for scientificreasoning, 68.98 on IFBench for instruction following, and 88.57 on AIME25 formathematical problem solving. Before optimization, the original model scored79.1, 67.3, and 92.8, respectively. Based on the simple average of the threemajor evaluation categories, the optimized model maintained approximately 99.2%of the original model’s accuracy.

This achievement goes beyond simply runninga large AI model on a Korean NPU. It confirms that performance and stabilityrequired for real-world services can be maintained in such an environment. Inparticular, the result is seen as a meaningful example of how FuriosaAI’s datacenter NPU, LG AI Research’s advanced AI capabilities, and Nota AI’s modeloptimization technology can work together to support the operation ofhigh-performance LLMs within Korea’s AI ecosystem.

Recently, access to cutting-edge AI modelsand the infrastructure required to run them has become an important issue inthe global AI industry. Following discussions around export controls on certainAI models and infrastructure, countries are increasingly focusing on securingtheir own AI models and computing infrastructure as part of the sovereign AImovement. In this context, Nota AI’s achievement highlights the need fordomestic AI semiconductors, domestic AI models, and the optimizationtechnologies that connect them to advance together.

“As sovereign AI gains increasingattention, what matters most is connecting models, semiconductors, andoptimization software into an executable AI infrastructure,” said Myungsu Chae,CEO of Nota AI. “This achievement demonstrates the real-world operationalpotential of large AI models by combining FuriosaAI’s data center NPU, LG’sflagship AI model K-EXAONE, and Nota AI’s optimization technology.”

‍