Upstage

Solar Open 2

바로 비교하기

바로 비교하기

Solar Open 2 is a Korean open-weight model developed by Upstage with 250 billion parameters and a strong focus on operating as an AI agent capable of completing real-world work. Rather than simply answering questions, it is designed to find relevant information, analyze multiple documents, use external tools, and carry complex tasks such as coding and document creation through to completion. Compared with the previous Solar Open model, it offers significant improvements across knowledge, mathematics, coding, instruction following, and Korean-language processing. In particular, it was trained through reinforcement learning and knowledge distillation using 12 teacher models specialized in areas such as coding and office work, enabling it to perform tasks reliably across a wide range of professional environments. Upstage also developed Office Verse, a simulation environment that recreates real office workflows. It generates large volumes of tasks such as spreadsheet analysis, document writing, and presentation creation, then automatically evaluates the resulting work to produce high-quality training data grounded in practical business scenarios. This strengthens the model’s ability to analyze information distributed across multiple documents and spreadsheets, identify problems during a task, and correct them independently. In a live demonstration, Solar Open 2 completed a commercial-area analysis task in approximately 30 minutes that would normally require about a week of human work. It also demonstrated strong practical performance in multi-step tasks such as legal review and the production of complex short-form videos. Solar Open 2 uses a hybrid architecture that combines Softmax Attention and Linear Attention at a ratio of three to one. This design reduces the memory usage and computation required for processing long contexts to approximately one-third of conventional approaches while maintaining high accuracy. With support for contexts exceeding one million tokens, the model can maintain continuity across large collections of documents, extensive codebases, and long-running workflows. It activates only the portion of its 250 billion parameters required for each task and applies quantization and pruning techniques, allowing a model of this scale to run on two H100 GPUs. The model supports Korean, English, and Japanese, with particularly strong Korean understanding and generation based on high-quality Korean training data. Because its weights are openly available, organizations can deploy it on their own internal servers or adapt it to proprietary data. This makes Solar Open 2 particularly suitable for businesses that cannot send sensitive information to external cloud services, as well as organizations building private AI agents and workflow automation systems.

AI 성능 평가