First Korean model on global developer platform

Agent performance sharply improved for real-world tasks

AAII score of 42 outpaces Big Tech rivals

A promotional image for Upstage's AI model Solar Pro 4 [Upstage]
A promotional image for Upstage's AI model Solar Pro 4 [Upstage]

Upstage has unveiled Solar Pro 4, a large language model with enhanced reasoning and AI agent capabilities, becoming the first Korean AI model listed on OpenRouter, a global developer platform.

The company said Friday it released Solar Pro 4, its next-generation commercial AI model. The new model improves on its predecessor, Solar Pro 3, in reasoning and agent performance, and is designed to reliably handle the complex workflows demanded in real-world work environments.

According to Upstage, Solar Pro 4 became the first Korean model to appear on a global developer platform immediately after its launch. It was listed on Hermes Agent, an AI agent developed by US-based Nous Research, and on OpenRouter, an API routing platform. Upstage said it is the first Korean company registered as an official provider on OpenRouter alongside major US technology companies and leading Chinese firms.

Developer uptake was swift. Solar Pro 4 surpassed 80 billion tokens in cumulative consumption within three days of its OpenRouter listing.

Upstage also said Solar Pro 4 became the first Korean model to enter the Agent Arena leaderboard, which compares agent performance across major global models. The model ranked on par with Nvidia's Nemotron 3 Ultra.

Upstage said it focused on improving "agentic tasks" in developing this model. Solar Pro 4 scored 57 on TerminalBench, a benchmark that evaluates a model's ability to complete multi-step tasks in a terminal from start to finish — 4.8 times higher than its predecessor. It also scored 23 on Tau3-Banking, a benchmark assessing tool-use ability through multi-turn dialogue, a 2.6-fold improvement.

Another key feature of Solar Pro 4 is its ability to handle long and complex documents. The model scored 71 on AA-LCR, a long-context comprehension benchmark that evaluates a model's ability to extract information from lengthy documents and reason toward an answer — more than 2.3 times better than its predecessor. Upstage said the model's reduced tendency to lose context in the middle of documents improved not only accuracy but also token efficiency.

Korean-language performance tailored to the office environments of domestic companies and public institutions has also improved. Solar Pro 4 is set to be integrated into Upstage Studio, the company's all-in-one agent document processing solution. The model can extract information from a wide range of document formats, including hangul documents, and handle tasks such as summarization, analysis and translation.

Solar Pro 4 scored 42 overall on the Artificial Analysis AI benchmark, more than three times the score of its predecessor. It outperformed Big Tech models including Nvidia's Nemotron 3 Ultra, which scored 38, and Google's Gemini 3.5 Flash-Lite at 37. It also posted a wide margin over competing sovereign models such as Mistral's Medium 3.5, which scored 30, and Cohere's Command A+ at 23.

"What AI needs in the real world is not the ability to ace tests, but the reliability to see a job through to the end," said Kim Sung-hoon, CEO of Upstage. "Solar Pro 4 becoming the first Korean model listed on a global developer platform is proof that our AI technology works on the world stage, and we will continue building AI that is even more practical and useful."


chami@heraldcorp.com