Upstage's Solar Pro 3 is the latest large language model developed by leveraging extensive operational experience and technical selections from real-world service environments. Building upon the foundation of Solar Pro 2, which established a standard for immediate applicability in enterprise settings due to its high efficiency and stability, Solar Pro 3 significantly elevates reasoning accuracy and instruction following quality while preserving the same API usability and processing speed. The core design principle of Solar Pro 3 revolves around enhancing response quality without compromising existing API interfaces, request methods, or throughput, allowing existing Solar Pro 2 users to experience immediate quality improvements simply by swapping the model.

From a performance perspective, Solar Pro 3 maintains a similar level of throughput (TPS) and cost structure as Solar Pro 2, but achieves meaningful performance gains in areas critical for user interaction. Specifically, it demonstrates a 52% improvement in instruction following capabilities (IFBench score of 55.78) and a 30% enhancement in complex reasoning tasks (Arena Hard v2 score of 62.5). Comparative benchmarks show consistent performance improvements in Instruction Following, Preference Alignment, and Reasoning, leading to more stable and accurate responses under the same processing speed conditions.

The development methodology for Solar Pro 3 was driven not by single benchmark optimization, but by addressing practical limitations frequently observed in real-world operating environments. The model focuses on improving response consistency and accuracy for:

1. Multi-step reasoning queries: Enhancing the ability to handle complex chains of thought.
2. Tasks with nuanced instructions: Accurately interpreting subtle differences in user directives that significantly impact outcomes.
3. Complex preference judgments: Delivering more reliable results for requests requiring intricate evaluation of preferences.

Significant improvements have been recorded in mathematical and logical reasoning, and instruction following metrics. This indicates a strengthened capacity for the model to accurately understand user requests and resolve them comprehensively. Despite its enhanced expressiveness as a large model, Solar Pro 3 maintains the serving efficiency and response speed of Solar Pro 2, thereby minimizing the typical trade-off where performance gains lead to increased costs. It also boasts a broader and deeper learning scope, improving response quality across general knowledge, specialized domain knowledge, and complex reasoning. A key objective of Solar Pro 3 is to more accurately reflect user intent and context in both Korean and English environments.

Solar Pro 3 is engineered for stability in practical applications within enterprise environments. It ensures compatibility with existing systems, predictable operational costs, and stable processing speeds, while simultaneously boosting reasoning accuracy and response quality. It is robustly designed for diverse enterprise scenarios, including complex query processing, multi-stage reasoning, and precise interpretation of user intentions.

Solar Pro 3 is currently available via API through the Upstage Console and can be accessed through other channels like yupp.ai and BizRouter. A free trial is offered until March 2, 2026, 23:59 UTC, to allow users to directly experience its improved reasoning quality in actual workloads.