Press Release
July 27, 2026
Nota AI (CEO Myungsu Chae), a company specializing in AI model light weighting and optimization, announced today that it has releaseda dedicated lightweight version of "Solar Open 2," the second-phase deliverable of the government-led sovereign AI foundation model project it is participating in alongside Upstage.
As high-performance open-weight models such as DeepSeek and Moonshot AI's Kimi series draw growing market attention, the basis of AI competition is expanding beyond raw performance to real-world agentic usability and operational cost efficiency. Nota AI's latest achievement is expected to significantly reduce the infrastructure burden of large-scale AImodels, accelerating enterprise AI transformation.
Solar Open 2 is a large language model designed forenterprise-grade agentic AI use cases, including document processing, coding,tool calling, and multi-step reasoning. The model has 250 billion total parameters but activates only 15 billion during inference, using aMixture-of-Experts (MoE) architecture. This approach selectively activates onlya subset of the model's computational modules at a time, improvingcomputational efficiency — particularly valuable for token-intensive agenticworkloads.
To maximize deployment efficiency for Solar Open 2 in real enterprise environments, Nota AI applied three core techniques:data-driven MoE quantization, router-aware MoE quantization, and non-uniform global pruning. For data-driven MoE quantization, the company used its proprietary technique, "PASCAL-MoE," which synthesizes calibration data to ensure all expert modules maintain stable statistical properties while converting model weights into 4-bit low-precision formats based on INT4 or NVFP4. Router-aware MoE quantization incorporates a proprietary quantization algorithm designed to preserve stable expert selection by the router even after quantization. Additionally, going a step beyond conventional quantization-centric optimization, Nota AI introduced non-uniform global pruning, which selectively removes lower-importance expert modules.
As a result, Nota AI successfully reduced the original model's weight memory from 500.6GB to 117.8GB — a reduction of approximately 76.5% — by converting model weights to 4-bit format and selectively pruning lower-importance computational modules. In terms of memory capacity, this optimization allows the model — which originally required 8 NVIDIA H100 GPUs —to run on just 2, while maintaining stable performance of tool calling, a core function for agentic AI, even after quantization.
This achievement goes beyond simply reducing model size, representing a level of optimization suitable for real enterprise deployment. Industry observers say the advance improves the practicality of enterprise AI adoption and helps reduce infrastructure costs throughout the AI Transformation (AX) process.
"This achievement is significant in that it optimizes a large-scale MoE model into a form that enterprises can realistically adopt," said Taeho Kim, CTO of Nota AI. "We will continue to advance quantization and light weighting technologies optimized for different model architectures and hardware environments, so that companies can adopt agentic AI and expand its use according to their own infrastructure and budgets."
The lightweight version of Solar Open 2 released by Nota AI is available on the global AI platform Hugging Face.
