Upstage said on the 14th that it unveiled Solar Pro 4, a large language model (LLM) that boosts reasoning and artificial intelligence (AI) agent performance. The model focuses on real work execution beyond simple Q&A, including long-form analysis and information extraction, tool use, and multi-step decision-making.
Solar Pro 4 became the first among domestic AI models to be adopted by Hermes Agent of Nouse Research in the United States and the global API brokerage platform OpenRouter. On OpenRouter, it surpassed a cumulative usage of 80 billion tokens just three days after listing. It also entered Agent Arena, which pits the agent capabilities of global models against each other, as the first Korean model and reached a performance tier similar to Nvidia's Nemotron 3 Ultra.
Work-execution metrics also improved. It scored 57 on TerminalBench v2.1, which measures the ability to handle multiple tasks in a terminal, recording a score 4.8 times higher than the previous version. Its score on Tau3-Banking, an evaluation of conversational tool use, rose 2.6 times to 23. In the long-form comprehension evaluation AA-LCR, it scored 71, exceeding the previous model by 2.3 times.
In the overall assessment, it scored 42 under the standards of Artificial Analysis, a global AI evaluation agency. The company said it outperformed Nemotron 3 Ultra (38) and Google Gemini 3.5 Flash-Light (37).
The new model will also be applied to the document processing solution Upstage Studio. Corporations and public institutions can continuously extract information from documents in various formats, including Korean, and carry out summarization, analysis, and translation.
Upstage CEO Kim Sung-hoon said, "The practical competitiveness of AI depends on the reliability of completing assigned tasks to the end," and added, "We will continue to introduce highly usable AI that is validated in the global market."