Nota, a company specializing in AI model optimization, has unveiled a lightweight model dedicated to Upstage's large-scale language model (LLM), 'Solar Open 2'. This technology focuses on improving the cost efficiency of enterprise AI transformation (AX) by reducing the infrastructure burden required to operate large-scale AI models.

Nota announced on the 27th that it has unveiled a lightweight model dedicated to 'Solar Open 2' as the second outcome of the government's independent AI Foundation model project, in which it is participating jointly with Upstage.

With the recent proliferation of high-performance openweight AI models such as DeepSeek and Moonshot AI, the criteria for AI competitiveness are expanding beyond just model performance to include actual business applicability and operational cost efficiency. Nota's strategy is to use this technology to support companies in operating large-scale AI models in a smaller infrastructure environment.

Solr Open2 is a large-scale language model with a total of 250 billion parameters, and adopts a Mixture of Experts (MoE) structure that activates only about 15 billion parameters when performing tasks.

Nota stated that it applied data-driven MoE quantization, router-aware MoE quantization, and non-uniform global pruning techniques for model optimization.

For data-driven MoE quantization, 'PASCAL-MoE' was utilized to convert model weights into a 4-bit (INT4·NVFP4) format, and a proprietary algorithm was applied to ensure the expert selection process remains stable even after quantization. Additionally, the model size was reduced by applying a non-uniform global pruning technique that selectively removes expert modules of low importance.

According to Nota, as a result of optimization, the weighted memory of Solar Open2 decreased from the existing 500.6GB to 117.8GB. This corresponds to a weight reduction effect of approximately 76.5%.

The company explained that it optimized the original model, which required eight NVIDIA H100 GPUs based on memory capacity, so that it could be operated in an environment with two H100 GPUs.

They also stated that they maintained the performance of Tool Calling, a key feature of Solar Open2, even while reducing the model capacity.

"It is significant that we have optimized a large-scale MoE model into a form that companies can actually implement," said Kim Tae-ho, CTO of Nota. "We plan to continuously advance our lightweight technology to enable the adoption of agent AI tailored to companies' infrastructure environments and budgets."

Nota unveiled the lightweight model dedicated to the Solar Open 2 through the global AI platform Hugging Face.