Solar Open2 250B

Solar Open2 250B is an open-weight large language model developed by Upstage AI, specifically engineered for agentic workflows, multi-step reasoning, and long-horizon tasks. Released in July 2026, it serves as a key component of South Korea's "sovereign AI" initiative and as a successor to the original Solar Open series. The model is optimized for real-world office productivity, tool calling, and coding across English, Korean, and Japanese.

The model utilizes a Hybrid-Attention Mixture-of-Experts (MoE) architecture, featuring a total of 250 billion parameters with a sparse activation of approximately 15 billion parameters per token. Its attention mechanism interleaves one standard softmax-attention layer with three linear-attention layers in every four-layer block. By removing rotary positional encoding (NoPE) and utilizing this hybrid stack, the model supports a massive context window of 1,048,576 tokens (1M) while maintaining significantly lower memory and compute costs than standard transformer architectures.

Solar Open2 250B was trained on a high-value 10-trillion-token dataset refined from a larger 20-trillion pool. To optimize training efficiency, Upstage initialized the model through selective weight transfer from the earlier 102B Solar Open model, which allowed for faster convergence. The model's agentic capabilities were further refined through Multi-teacher On-Policy Distillation (MOPD), a technique used to consolidate knowledge and skills from twelve specialized domain models into a single, cohesive foundation model.

Distributed under the Upstage Solar License, the model is available for commercial use and modification, provided that derivative models include the "Solar" prefix in their names and provide appropriate attribution. Due to its efficient MoE design and support for advanced quantization formats like NVFP4, the model is capable of running on enterprise-grade hardware, such as dual NVIDIA H200 configurations, making it accessible for localized corporate and institutional deployments.