[Edaily Reporter Yoon Jeong-hoon] #NOTA, a company specializing in AI model lightweighting and optimization technologies, has once again demonstrated its technological competitiveness at the world’s most prestigious machine learning conference.

NOTA announced on the 11th that two of its papers on MoE (Mixture-of-Experts) specialized quantization algorithms have been accepted for the “Resource-Adaptive Foundation Model Inference (AdaptFM)” workshop at the upcoming “ICML 2026 (International Conference on Machine Learning).”

ICML is one of the world’s leading conferences in the fields of machine learning and artificial intelligence, serving as a platform where the latest AI research achievements from global tech giants, major universities, and research institutions are presented. The AdaptFM workshop, where these papers were accepted, focuses on technologies that enable the efficient execution of large-scale AI models even with limited computing resources. Researchers from global companies such as Amazon and Meta, as well as major research institutions, are participating in the organizing committee, while researchers from leading AI companies including NVIDIA, Qualcomm AI Research, OpenAI, Apple, and Microsoft are serving as program committee members.

This achievement is significant as it recognizes the technical expertise NOTA has accumulated in the field of MoE (Mixture-of-Experts) model optimization, a core architecture for large language models that has recently garnered significant attention. MoE operates by selecting only the necessary components from a set of expert models, enabling simultaneous improvements in both performance and efficiency for large AI models, which is why it is rapidly gaining traction in the latest LLMs. However, given the complexity of the model architecture, a different approach is required during the quantization process—which aims to make the model smaller and lighter—compared to conventional models.

Following its track and overall victories at the recent NVIDIA Nemotron Hackathon with a data-driven MoE quantization technique, NOTA is now presenting research results specialized for MoE architectures at this workshop.

The first paper selected for this workshop, “DREAM-MoE,” proposes a method to reduce changes in decision flows that can occur when quantizing large-scale AI models by dividing them into multiple segments. Noting that a small error in an earlier segment can alter expert selection in later segments, the paper ensures that the model can continue to select the necessary experts in a manner similar to the original even after quantization.

The second paper, “SRA-MoE,” proposes a method to identify and prioritize the protection of critical inputs that have a greater impact on model outcomes. Rather than treating all inputs equally, the design ensures that expert selection remains stable for key inputs, enabling effective maintenance of model quality even with limited resources.

Both studies demonstrated higher performance compared to the latest MoE-specific quantization techniques. This shows that large-scale AI models can be run with less memory and computational resources while minimizing quality degradation. As the costs, power consumption, and hardware burdens associated with operating large AI models continue to rise, the importance of MoE-specific quantization technology is growing.

NOTA is proactively focusing its R&D capabilities on the field of optimizing large AI models that require significant memory and computational resources. While advancing the optimization of large-scale models such as Solar MoE through the Upstage Consortium’s proprietary foundation model project, the company is also expanding its technology’s scope by extending its experience in quantizing NVIDIA Nemotron 3 nano models to the latest large models, such as Nemotron Ultra.

Chae Myeong-soo, CEO of Nota, stated, “The acceptance of this paper is a testament to Nota’s consistent efforts to advance quantization technology specialized for MoE.” He added, “Following our overall victory at the NVIDIA Nemotron Hackathon, we are now presenting our research findings at the ICML 2026 AdaptFM Workshop. We will continue to develop optimization technologies that enable more efficient utilization of large-scale AI models.”

Meanwhile, Nota plans to host “Nota AI - Korea Efficient Days” during ICML 2026, which will take place at COEX in Samseong-dong, Seoul, starting on the 6th of next month. The event aims to share research trends and industrial application possibilities of Efficient AI with global researchers, engineers, and corporate representatives visiting Korea. Through this event, Nota plans to introduce its research achievements in the field of large-scale AI model optimization and expand opportunities for technological collaboration and business partnerships.