Upstage CEO Kim Sung-hoon speaks during his presentation at a media conference at Conrad Hotel Seoul, June 16. Yonhap

Upstage, a Korean artificial intelligence (AI) company, said Thursday it has released Solar Open 2, an open-weight large language model designed to complete multistep tasks rather than simply answer questions.

The company published the model's weights, the underlying parameters that determine how the AI processes information, on Hugging Face, a global open-source platform, along with a technical report detailing its architecture and training methods.

The model was developed under a government-backed project to build Korea's own foundation AI models, in which Upstage serves as the lead organization.

Solar Open 2 has 250 billion parameters but activates only 15 billion at a time, using a "mixture of experts" structure that Upstage said makes it more efficient for token-heavy agent tasks. It can process up to 1 million tokens of context and, when compressed through quantization, can run on two Nvidia H200 chips.

The company said the model outperformed DeepSeek's V4 Flash, Mistral's Medium 3.5 and Cohere's Command A+ on benchmarks measuring instruction-following, tool use and overall agent performance, and improved more than 20 points over its predecessor on reasoning, math and coding tests. In Korean-language benchmarks, it also outperformed OpenAI's GPT-5.4 mini and Anthropic's Claude Haiku 4.5.

The model is released under the Apache 2.0-based Upstage Solar License, permitting commercial use and derivative development.

Upstage plans to apply it to the Daum portal and its Tamnlee AI agent platform, and to expand its use across finance, law, health care and other industries alongside domestically made AI chips.

"This model was built around the usability of an agent that completes work on its own, not just benchmark scores," Upstage CEO Kim Sung-hoon said.

This article was published with the assistance of generative AI and edited by The Korea Times.