Solar Pro 3 is Upstage's large language model designed around a Mixture-of-Experts architecture with 102B total parameters and only 12B active per forward pass, delivering capable performance while keeping compute and serving costs predictable. The model is optimized for Korean while still supporting English and Japanese, and Upstage's March update emphasized raising Korean-language quality to reasoning-level parity with its English output, a notable differentiator for teams running agentic workflows in Korean-speaking environments. Underneath, improvements come from SnapPO, Upstage's proprietary reinforcement learning framework developed through the Solar Open research program, which is structured so each training stage can run and compose independently across math, code, and agentic domains.
Solar Pro 3 was positioned as a meaningful step up from Solar Pro 2, roughly doubling scores across core agentic benchmarks while keeping the same API surface and serving behavior. Upstage reports Tau2-all rising to 72.3 from 36.0, SWE Bench to 28.6 from 14.5, and Terminal Bench 2 to 10.1 from 2.2, gains that the lab attributes to design decisions informed by real production failures rather than narrow benchmark tuning. Independent evaluation through Artificial Analysis places the model at 14 on the Intelligence Index, with strengths in conversational agents (τ²-Bench Telecom at 86.3%) and graduate-level scientific reasoning (GPQA Diamond at 72.4%), making it a practical fit for tool-using pipelines that need reliable multi-step planning, precise instruction following, and natural Korean responses without the usual English-quality gap.
