Upstage unveiled its in-house large language model (LLM) "Solar Pro 3" on the 24th. As the center of gravity in the recent artificial intelligence (AI) race shifts beyond raw model performance to Agentic AI that performs real work, Upstage has launched a full-fledged push for practicality with its new model.
Solar Pro 3 has 102 billion parameters, more than triple that of the previous "Solar Pro 2." Even so, the company said it boosted efficiency by keeping expense and throughput (TPS) at the same level as its predecessor. In performance, it delivered meaningful gains across the agent workflow, including tool calls for multi-step task execution and adherence to complex instructions. Upstage said it more than doubled scores over the previous model on major benchmarks, including Tau2-all for overall agent performance, Terminal Bench 2 and SWE Bench for coding, and IFBench for instruction following.
Upstage also raised deep reasoning capabilities by applying its in-house Reinforcement Learning technology "SnapPO." It emphasized that, as a result, it achieved a performance leap on competition-level math problems such as HMMT '26 and AIME '26, and on graduate-level science evaluations such as GPQA-Diamond. It also logged improvements on Arena-Hard-v2, a general user preference metric, and Ko-Arena-Hard-v2, a Korean user preference metric.
Solar Pro 3 is available immediately through the global AI model service platform OpenRouter and Upstage's own API. Upstage CEO Kim Sung-hoon said, "We developed it with the goal of a leap in the practicality of agent AI that delivers results in real work environments," adding, "We will expand the AI ecosystem used on industrial sites."