Upstage: Solar Pro 3

Solar Pro 3 is a Mixture-of-Experts language model built by South Korean AI developer Upstage, designed around a sparse activation architecture that keeps operational costs and processing speeds close to its smaller predecessor despite scaling to 102 billion total parameters. With 12 billion active parameters engaged per forward pass, the model prioritizes computational efficiency while delivering the depth needed for complex, multi-step reasoning. Its 128,000-token context window allows it to handle extended documents, long conversations, and involved workflows that demand sustained coherence across large amounts of information. The model is particularly optimized for Korean-language tasks while extending strong support to English and Japanese, positioning it as a practical choice for multilingual applications in the Asia-Pacific region and beyond.

Upstage developed SnapPO, a proprietary reinforcement learning technique, to refine the model's reasoning capabilities and push performance gains across agentic workflows—achieving roughly double the agentic AI capability, coding proficiency, and instruction-following performance seen in its predecessor. The focus on agentic scenarios means Solar Pro 3 is built to handle tasks that require planning, tool use, and multi-step decision-making rather than single-pass generation. The model is immediately accessible through OpenRouter and Upstage's own API, making it straightforward to integrate into agents, copilots, and automated pipelines where reliable tool-calling and structured output generation matter. The combination of scalable architecture, specialized training signal, and practical deployment options makes Solar Pro 3 a strong candidate for teams building AI systems that operate independently across complex environments.