Nota AI Advances to Round 3 of Korea's Sovereign AI Foundation Model Project Alongside Upstage

- Nota AI leads model compression and optimization within the consortium, reducing GPU requirements from eight NVIDIA H100s to two
- Lightweight version of Solar Open 2 surpasses 80,000 downloads on Hugging Face, reflecting strong demand for deployable, resource-efficient models

‍

Nota AI, a company specializing in AI model optimization and compression technology, announced today that it will continue providing technical support to help Korea's sovereign AI model run efficiently across diverse service environments, following the advancement of the Upstage-led consortium to Round 3 of the government's Sovereign AI Foundation Model project.

Within the Upstage consortium, Nota AI is responsible for compressing and optimizing the Solar foundation model family so it can run efficiently within domestic infrastructure environments. The company's role is to reduce the computing resources and memory footprint required to run the model while preserving its performance, enabling deployment across a wider range of hardware and service environments. This work is intended to lower the cost and operational burden of adopting Korea's national AI model and support its broader use across domestic industry.

Ahead of the announcement of Round 2 evaluation results for the Sovereign AI Foundation Model project, Nota AI released a lightweight version of Upstage's Solar Open 2 foundation model on Hugging Face, the global AI model-sharing platform, to demonstrate this optimization capability. The lightweight version reduced the infrastructure required to run the model from eight NVIDIA H100 GPUs to two, lowering deployment costs and expanding the model's potential for wider adoption. The lightweight model has since surpassed 80,000 cumulative downloads, underscoring strong interest in practical, deployable compressed models.

Solar Open 2, developed by Upstage, is a frontier-class foundation model with 250 billion parameters. Nota AI applied its optimization technology to convert the model's weights into 4-bit formats based on INT4 or NVFP4, while selectively pruning lower-importance computational modules.

As a result, the memory required to store the model's weights was reduced from 500.6GB to 117.8GB, a 76.5% reduction. Based on memory requirements, this reduction made it possible to run the model on two NVIDIA H100 GPUs, down from the eight GPUs required for the original model.

"We will continue to advance our optimization technology so that Korea's national AI model can be used cost-effectively across a wide range of industries," said Myungsu Chae, CEO of Nota AI. "Together with Upstage, we will do our utmost to achieve strong results in the Round 3 evaluation of the Sovereign AI Foundation Model project."