Solar Pro 2

31B-parameter proprietary model with 64K context, continuing the Solar line's depth-up-scaling lineage. Upstage frames it as an upgrade of Solar Pro (22B, itself depth-up-scaled from Phi-3-medium 14B); ZDNet Korea's 2025 sovereign-AI tier taxonomy classifies it as a fresh DUS of Phi-4 (14B). Either way the weight lineage is external (Microsoft Phi): Upstage published no tech report for it, never claimed from-scratch pretraining, and later conceded its pre-Solar-Open models "were redesigned from reference models using DUS" — reserving "from scratch" for the Solar Open line. Two operating modes: Chat (fast, fluent) and Reasoning (multi-step CoT for math, code, logic). Designed to fit on a single GPU.

AA Intelligence Index: 8 (non-reasoning). Matches or exceeds GPT-4o, DeepSeek R1, and Qwen 3 on several benchmarks despite being less than half their size. Leads all models on Korean benchmarks (Ko-MMLU, Hae-Rae, Ko-IFEval).