Upstage unveiled its self-developed large language model (LLM) Solar Pro 4 on Friday.
Solar Pro 4 is Upstage’s next-generation commercial artificial intelligence (AI) model, boasting enhanced inference capabilities and agent performance compared to its predecessor, Solar Pro 3.
The company stated that the model is optimized for real-world applications rather than single-answer benchmark scores. It excels in document comprehension, information extraction, long-context reasoning, and continuous decision-making – tasks crucial in actual work environments.
Shortly after its release, Solar Pro 4 made history as the first Korean model to be featured on the OpenRouter API routing platform and Hermes Agent, an AI agent developed by U.S.-based Nuance Research.
This milestone marks the first time a Korean company has been officially listed as a provider alongside major U.S. tech giants like OpenAI, Anthropic, Google, and NVIDIA, as well as leading Chinese AI models such as DeepSeek, Moonshot, and Minimax.
Additionally, it entered the Agent Arena leaderboard, which compares the agent performance of major global models, achieving comparable results to NVIDIA’s Nemotron 3 Ultra – another first for a Korean model.
Within just 72 hours of its OpenRouter listing, Solar Pro 4 processed over 80 billion tokens.
Upstage highlighted the model’s proficiency in handling lengthy and complex documents, a critical factor in determining the success of agent tasks.
It scored 57 points on the Terminal-Bench v2.1, which evaluates multi-step task completion abilities, marking a 4.8-fold improvement over its predecessor. In the τ³-Banking assessment, which measures tool usage through conversation, it achieved a score of 23, representing a 2.6-fold enhancement.
The model demonstrated exceptional performance in the long document comprehension benchmark (AA-LCR), scoring 71 points – more than 2.3 times better than the previous version. This test evaluates the ability to extract information from extensive documents and infer answers based on that data.
Overall, Solar Pro 4 achieved a score of 42 according to Artificial Analysis, a global AI evaluation agency, placing it among top-tier general-purpose frontier models.
It outperformed big tech models such as NVIDIA’s Nemotron 3 Ultra (38 points) and Google’s Gemini 3.5 Flash-Light (37 points), while also surpassing competitive sovereign models like Mistral’s Medium 3.5 (30 points) and Cohere’s Command A+ (23 points).
Kim Seong-hoon, Chief Executive Officer (CEO) of Upstage, emphasized that in practical applications, AI’s most crucial attribute is its reliability in completing assigned tasks. The model’s inclusion on a global developer platform as the first Korean entry proves that the AI technology is competitive on the world stage.
