A Korean research consortium led by KAIST and Upstage published Solar Open 2, a 250-billion-parameter mixture-of-experts language model built specifically for extended AI agent workflows.

The model scales up from Solar Open 1's 100-billion parameters and targets what the researchers call "long-horizon agentic tasks" — multi-step AI operations that require maintaining context across lengthy interaction sequences.

Technical Architecture

Solar Open 2 achieves its one-million-token context window through a hybrid attention mechanism that alternates between standard softmax attention and linear attention layers. The architecture uses no positional encoding and implements what the team calls a "gated delta rule extended to negative eigenvalues."

The researchers initialized the new model from Solar Open 1, transferring 5.69 billion parameters from the shared architectural components while training the remainder from scratch. This approach allowed them to work within a fixed compute budget while scaling up model size.

For training data, the team refined a 20-trillion-token dataset down to 10 trillion tokens through what they describe as "quality- and rarity-aware data curation." The curated mixture outperformed the original Solar Open 1 training recipe at equivalent token budgets.

Agent Training Process

The model's agent capabilities come from training twelve domain-specific expert models across purpose-built scenarios, then consolidating them into a single model through Multi-teacher On-Policy Distillation (MOPD).

On English benchmarks, Solar Open 2 leads on MMLU-Pro, LiveCodeBench, and the APEX-Agents agentic evaluation suite. It remains competitive with DeepSeek-V4-Flash and MiMo-V2.5 on other standard tests.

For Korean language tasks, the model achieved the highest average scores compared to any model tested, including closed commercial APIs. On Ko-GDPval, an internal Korean office-work agent benchmark, it matched DeepSeek-V4-Pro's performance while using less than one-sixth the parameters.

The 53-author paper represents collaboration between KAIST, Upstage, and multiple Korean research institutions. The model targets applications requiring sustained reasoning across extended contexts, particularly in enterprise automation scenarios.