[Edaily Reporter Yun Junghoon ] #Nota Inc., a company specializing in AI model lightweighting and optimization technologies, announced that it has successfully optimized LG Corp.’s national flagship AI model, K-EXAONE 236B, on FuriosaAI’s data center-grade NPU.

K-EXAONE 236B is a large-scale AI model with approximately 236 billion parameters that adopts a MoE (Model of Experts) architecture, which selectively utilizes multiple expert models. While the MoE architecture can improve the efficiency of large models, the optimization process requires sophisticated technology to ensure that each expert model operates stably.

In particular, frontier-class large models undergo lengthy inference processes when solving complex problems, so even minor errors arising during quantization can accumulate and affect the accuracy of the final answer. This achievement is significant because it demonstrates the ability to optimize such large models for a domestically developed NPU environment while maintaining accuracy in key evaluations.

In this project, Nota Inc. optimized K-ExaOne for FuriosaAI’s data center NPU environment. Rather than re-training the entire model, the team precisely analyzed specific sections prone to performance degradation and applied optimizations only where necessary, thereby minimizing performance loss. As a result, the large AI model ran efficiently on a domestic NPU environment while maintaining performance on par with the baseline model across key performance metrics.

Meaningful results were also confirmed in performance evaluations. Nota Inc. reduced the size of the K-ExaOne model by approximately 71%, lowering the memory burden required to run large AI models, while maintaining accuracy levels similar to the original model across key evaluation categories such as scientific inference, instruction comprehension, and math problem-solving.

This indicates that the model—which has the same scale of 236 billion parameters—has been optimized for more efficient execution, demonstrating the potential to enhance the operational efficiency of data center AI infrastructure.

Based on Nota Inc.’s internal evaluation environment, the optimized model scored 79.80 on the scientific reasoning task (GPQA), 68.98 on the instruction understanding task (IFBench), and 88.57 on the math problem-solving task (AIME25). The performance of the original model before size reduction was 79.1, 67.3, and 92.8, respectively; even after optimization, it maintained approximately 99.2% of the original’s accuracy based on the simple average of the three key evaluation metrics.

This achievement is significant because it demonstrates that large AI models can not only be run on domestically developed NPUs but can also maintain the performance and stability required for actual service deployment. In particular, this is regarded as a case that demonstrates the feasibility of operating high-performance LLMs within the domestic AI ecosystem through the combination of FuriosaAI’s data center NPUs, LG Corp.’s advanced AI technology, and Nota Inc.’s AI model optimization technology.

In the global AI industry, access to cutting-edge AI models and the infrastructure that powers them has recently emerged as a critical issue. In particular, following discussions on export controls surrounding certain AI models and infrastructure, the “Sovereign AI” trend—in which nations seek to secure their own domestic AI models and computing infrastructure—is also gaining attention. Against this backdrop, this achievement underscores the need for domestic AI semiconductors, domestic AI models, and the optimization technologies that connect them to advance in tandem.

Chae Myeong-soo, CEO of Nota Inc., stated, “Amid the growing focus on sovereign AI, what matters most is connecting models, semiconductors, and optimization software into a single, viable AI infrastructure.” He added, “This achievement is a case where FuriosaAI’s data center NPU, LG Corp.’s national AI model K-ExaOne, and Nota Inc.’s optimization technology were combined to confirm the practical operability of large-scale AI models.”