AI startup Upstage unveiled its in-house large language model, Solar Pro 3, on Monday.
A key feature is its 102 billion parameters, more than tripling the size of its predecessor, Solar Pro 2. The company said it achieved more than a twofold improvement over the previous version on major benchmarks including overall agent performance (Tau2-all), coding (Terminal Bench 2 and SWE Bench) and instruction following (IFBench). It kept costs and processing speed (TPS) the same as the previous version.
It also improved reasoning. It applied its in-house reinforcement learning technology, SnapPO, to advance deep reasoning based on step-by-step thinking. Performance gains were also confirmed in high-difficulty reasoning evaluations including competition-level math (HMMT'26 and AIME'26), graduate-level science (GPQA-Diamond) and Korean user preference (Ko-Arena-hard-v2).
Solar Pro 3 is available immediately through the global AI model platform OpenRouter and the Upstage API.
Upstage CEO Sung-hoon Kim (김성훈) said, "We developed it with the goal of a leap in practical agentic AI that can deliver results in real work environments."
Seulgi Son
sageson@d-today.co.kr