Upstage said on Thursday it has unveiled Solar Pro 4, a large language model with strengthened agent functions.

Solar Pro 4 is a commercial AI model that improves reasoning and agent performance over its predecessor Solar Pro 3. It strengthened its ability to perform tasks used in real work environments, including document understanding and information extraction, long-context reasoning and sequential decision-making.

Upstage said Solar Pro 4 became the first Korean model to be listed on the U.S. Nous Research AI agent Hermes Agent and the API routing platform OpenRouter. Cumulative token usage surpassed 80 billion within three days of its listing on OpenRouter.

It also became the first Korean model to be listed on the Agent Arena leaderboard, which compares agent performance among major global models. The company said it recorded a similar level of performance to Nvidia's Nemotron 3 Ultra on that leaderboard.

By benchmark, it scored 57 on Terminal-Bench v2.1, which evaluates the ability to perform multi-step tasks in a terminal. That is 4.8 times higher than its predecessor. It scored 23 on tau^3-Banking, which evaluates tool-use capability during conversations, an improvement of 2.6 times from its predecessor.

On AA-LCR, a long-document comprehension benchmark that evaluates the ability to extract information from long documents and reason based on it, it scored 71, more than 2.3 times higher than its predecessor.

Solar Pro 4 has also been applied to Upstage Studio, Upstage's agent document-processing solution. It can extract information from a range of documents, including Korean-language documents, and handle tasks such as summarisation, analysis and translation.

It scored 42 in the overall evaluation by global AI model assessor Artificial Analysis. Upstage said that was higher than Nvidia's Nemotron 3 Ultra (38), Google's Gemini 3.5 Flash-Lite (37), Mistral's Medium 3.5 (30) and Cohere's Command A+ (23).

Upstage CEO Seonghun Kim (김성훈) said, "What AI needs in practical work is not the ability to do well on tests, but the reliability to carry out assigned tasks through to the end." He added, "We will continue to improve the performance of AI models that can be used in the field."