Artificial intelligence (AI) model slimming and optimization corporations Nota unveiled on the 27th a dedicated slimming model for "Solar Open 2," the second-phase output of the government independent AI foundation model project it is undertaking with Upstage.
Solar Open 2 is a large language model (LLM) focused on implementing agent AI that can be used for real corporate work such as document tasks and coding, tool calling, and multi-step reasoning. It has a total of 250 billion parameters, and for actual tasks it activates only 15 billion as a "mixture of experts (MoE)" architecture. It boosts computational efficiency by selecting and running only some of the multiple computation modules within a single model.
Nota said it converted the model weights into a 4-bit format and selectively pruned lower-importance computation modules, cutting the original model's weight memory by about 76.5%, from 500.6 gigabytes (GB) to 117.8 GB. Based on memory capacity, it optimized the original model, which had required eight Nvidia H100 graphics processing units (GPUs), to run even in a two-GPU environment.
Nota Chief Technology Officer (CTO) Kim Tae-ho said, "This outcome is meaningful in that we optimized a large-scale MoE model into a form that corporations can actually adopt," adding, "We will continue to advance quantization and slimming technologies optimized for the model architecture and hardware environment so that corporations can introduce agent AI to fit their own infrastructure and budgets and expand the scope of its use."
The dedicated slimming model for Solar Open 2 can be found on the open-source AI platform Hugging Face.