Upstage has open-sourced Solar Open 2 on Hugging Face, bringing an efficient AI agent LLM from South Korea’s sovereign AI initiative with lower hardware requirements and competitive benchmark performance.
Upstage has released Solar Open 2, an open-source large language model (LLM) optimised for AI agent tasks, publishing its model weights on Hugging Face. Developed under South Korea’s government-backed “sovereign AI foundation model” initiative, with Upstage serving as the lead organisation, Solar Open 2 succeeds the original Solar Open released in January.
The model adopts a Mixture of Experts (MoE) architecture with 250 billion total parameters while activating only 15 billion during inference, reducing compute requirements for AI agent workloads involving repeated reasoning and tool calls. It supports context windows of up to one million tokens and, using quantisation, can run on just two NVIDIA H200 GPUs.
According to Upstage, Solar Open 2 outperformed DeepSeek V4 Flash and Mistral Medium 3.5 in agent-related benchmarks, including IFBench for instruction following and MCP-Atlas for tool calling. In the Ko-GDPval benchmark for South Korean industrial tasks, it scored 86.75, with the company claiming performance comparable to DeepSeek V4 Pro despite the latter being more than six times larger.
“While some recently released models require 16 NVIDIA B200 GPUs, Solar Open 2 is designed to operate with only two,” said Kim Sung-hun, CEO of Upstage.
Upstage plans to deploy Solar Open 2 on Daum, which serves more than 10 million weekly users, expanding from search and summarisation to conversational AI agent services. It will also support local governments and public institutions through its Timely AI agent platform while building AI infrastructure for finance, legal and healthcare using domestic NPUs.
