Upstage Releases Solar Open 2 — 250B Open-Weight AI Agent Model That Runs on 2 GPUs

South Korea's Upstage open-sources Solar Open 2, a Mixture-of-Experts LLM with 250B parameters that activates only 15B during inference, outperforming DeepSeek V4 Flash on agent benchmarks while running on just two H200 GPUs.

South Korean AI company Upstage has released Solar Open 2, an open-weight large language model purpose-built for autonomous AI agent tasks, publishing the model weights on Hugging Face under the country's government-backed sovereign AI initiative.

The model uses a Mixture of Experts (MoE) architecture with 250 billion total parameters — but activates only 15 billion during inference. That sparse activation keeps compute costs low for agent workloads, which involve repeated reasoning loops and tool calls. With quantization, Solar Open 2 can run on just two NVIDIA H200 GPUs, and it supports context windows of up to one million tokens.

On benchmarks, Upstage reports Solar Open 2 outperformed DeepSeek V4 Flash and Mistral Medium 3.5 on agent-specific evaluations including IFBench for instruction following and MCP-Atlas for tool calling. On Ko-GDPval, a benchmark for South Korean industrial tasks, it scored 86.75 — performance the company claims is comparable to DeepSeek V4 Pro, a model more than six times larger.

Upstage plans to deploy Solar Open 2 on Daum, South Korea's second-largest web portal with over 10 million weekly users, expanding from search and summarization into conversational AI agent services. The model will also support local governments and public institutions through the company's Timely AI agent platform, with additional infrastructure planned for finance, legal, and healthcare sectors running on domestic NPUs.

Sources: Open Source For You, Upstage