As demand for running Large Language Models (LLMs) on field devices or in limited GPU environments increases, competition for 'quantization' technology that reduces memory usage without significantly compromising model accuracy is intensifying.

Nota announced on March 5 that it had applied its quantization techniques to Upstage's 'Solar-Open-100B' to reduce memory usage.

According to the model card released by Nota on Hugging Face, the model weight memory footprint of Solar-Open-100B with 'Nota MoE Quantization' applied was presented as 51.9GB, down from 191.2GB. The same page explained that this approach is designed to mitigate **distortion** that can occur in the process of mixing experts in MoE (Mixture of Experts) structure-based LLM.

**Perplexity (PPL)** was also presented as a performance indicator. The table on the model card shows the PPL based on WikiText-2 as 6.06 for the original Solar-Open-100B and 6.81 for the Nota technique-applied model. PPL is used to indicate that the lower the value, the more stable the language model's predictions are.

This disclosure is noteworthy because, amid the proliferation of MoE-affiliated LLMs, it goes beyond a "universal reduction of the entire model" approach and instead presents a quantization approach that considers structural characteristics as an indicator of actual model card performance. However, some media reports have also highlighted the Ministry of Science and ICT's "Independent AI Foundation Model" project and whether a patent has been filed, and the extent of verification based on articles and publicly available data may vary.