Upstage, a South Korean AI startup, announced Thursday that it has released its internally developed large language model (LLM) "Solar Pro3."
Solar Pro3 is a large-scale artificial intelligence model with 102 billion parameters, more than three times the size of its predecessor Solar Pro2. Despite the increased model size, Upstage maintained the same cost and processing speed (TPS) as the previous model, enhancing efficiency.
In terms of performance, Solar Pro3 showed improvements across the entire agent workflow, including tool calling for multi-step tasks and complex instruction execution. Specifically, the model achieved more than double the performance of its predecessor in major LLM benchmarks, including overall agent performance (Tau2-all), coding (Terminal Bench 2 and SWE Bench), and instruction following (IFBench).
Upstage also applied its proprietary reinforcement learning technology "SnapPO" to advance the deep reasoning capabilities that are central to AI agents. Going beyond simple answer generation, the model employs step-by-step reasoning to improve consistency and contextual judgment. As a result, it achieved high scores in challenging reasoning evaluations such as competition-level math problem solving (HMMT'26 and AIME'26) and graduate-level science assessments (GPQA-Diamond).
Response quality also improved significantly. The model recorded meaningful gains in key evaluation metrics including general user preference (Arena-hard-v2) and Korean-language user preference (Ko-Arena-hard-v2). By accurately reflecting user intent and subtle nuances, Solar Pro3 enhanced the perceived quality in real-world usage environments. The model is available through OpenRouter, a global AI model service platform, and Upstage's own application programming interface (API).
"Solar Pro3 was developed with the goal of advancing AI agent practicality that can deliver results in real work environments, going beyond mere model performance," Upstage CEO Kim Sung-hoon said.