Upstage Open-Sources Solar Open 2: A 250B-Parameter MoE Model Purpose-Built for AI Agents

South Korean AI company Upstage released Solar Open 2 (250B-A15B) as an open-weight model on July 23, under an Apache 2.0-based license that permits commercial use and derivative development. The model is part of a Korean government-backed initiative to build the country's own foundation AI models, and is specifically designed for agentic use cases — tool calling, multi-step reasoning, and end-to-end task execution.

Solar Open 2 uses a mixture-of-experts architecture with 321 total experts (320 routed plus one shared), activating the top 8 per token for 15 billion active parameters out of 250 billion total. Its core architectural innovation is hybrid-attention: three linear-attention layers interleaved with one softmax-attention layer, repeated 12 times for 48 layers total. The model uses NoPE (No Positional Encoding), with linear attention encoding position through its recurrent state, avoiding the extrapolation limits of RoPE at long contexts. The model supports up to 1 million tokens of context.

On benchmarks, Solar Open 2 delivers competitive results. It scores 86.2 on MMLU-Pro, 86.3 on GPQA-Diamond, and 95.7 on AIME 2026. More notably, it achieves 16.6 on the APEX-Agents benchmark — well ahead of DeepSeek V4 Flash (13.2) and Mistral Medium 3.5 (13.4). It also outperformed OpenAI GPT-5.4 mini and Anthropic Claude Haiku 4.5 on Korean-language benchmarks. The model was pre-trained on approximately 12 trillion tokens using NVIDIA B200 GPUs, consuming roughly 2 million GPU hours.

A quantized version of Solar Open 2 can run on two NVIDIA H200 GPUs, with eight H200 GPUs recommended for full performance. Because only 12 of 48 layers maintain a KV cache, memory usage during inference is roughly one-quarter of an all-softmax model of comparable size. Upstage says the model was "built around the usability of an agent that completes work on its own, not just benchmark scores."

The model weights and a technical report are available on Hugging Face. Upstage plans to deploy Solar Open 2 across its Daum portal, its Tamnlee AI agent platform, and expand into finance, legal, and healthcare sectors alongside domestically produced Korean AI chips. For the open-source community, the model's standout agent benchmark scores and efficient MoE design make it a significant new option in the competitive open-weight LLM landscape.

Looking ahead, key questions include how Solar Open 2 performs in real-world agent workflows, whether the community builds a robust tool ecosystem around it, and whether the model helps establish South Korea's position in the global foundation model race.