Solar Open 2: Korea's Sovereign Foundation Model, Built for Agentic Use (www.upstage.ai)

Solar Open 2 has been announced as an open-weight foundation model optimized for agentic use, specifically designed to operate in real work environments. This model tackles a variety of agent tasks—such as coding, tool calling, and office work—while competing effectively with top global models based on multiple benchmarks across agent, Korean, knowledge, and coding tasks. A significant feature is its Mixture of Experts (MoE) architecture, activating only 15 billion of its total 250 billion parameters per token, which enhances both computational efficiency and performance. It supports a context window of up to 1 million tokens and runs efficiently on just two NVIDIA H200 GPUs when quantized, making it practical for enterprise deployments. The importance of Solar Open 2 lies in its shift from traditional language model functions to a more dynamic agent framework capable of executing long-horizon tasks and maintaining extensive context. This is achieved through specialized training data that reflects real work scenarios, enabling the model to perform reliably across various tasks. Additionally, its multilingual capabilities—officially supporting Korean, English, and Japanese—enhance its utility in diverse workplaces. Released on Hugging Face under a commercially usable license, it opens new avenues for development in AI-driven agent operations, showcasing significant improvements over its predecessor, Solar Open 100B, in performance and operational efficiency.