Solar Pro 3 is a 102-billion parameter large language model (LLM) developed by Upstage AI, officially unveiled on March 24, 2026. Designed as a specialized solution for Agentic AI, the model is engineered to handle multi-step tool calls, complex reasoning, and autonomous workflow execution. It represents a significant advancement over its predecessor, Solar Pro 2, tripling the total parameter count while utilizing architectural optimizations to maintain high throughput and cost-efficiency.
The model utilizes a Mixture-of-Experts (MoE) architecture, where only 12 billion parameters are active during any single forward pass. This configuration allows the system to deliver the reasoning capabilities of a 100B-scale model while operating with the speed and resource requirements of a much smaller system. It further incorporates Upstage's proprietary reinforcement learning technology, SnapPO, which is used to refine deep reasoning and instruction adherence, particularly for tasks involving graduate-level science and competition-level mathematics.
Performance and Capabilities
Solar Pro 3 is optimized for enterprise-grade applications, demonstrating high scores in benchmarks such as SWE Bench for coding and IFBench for instruction following. According to the developers, it has achieved performance gains of over 100% in overall agentic benchmarks (Tau2-all) compared to previous versions. Its reasoning capability is notably highlighted by its performance on advanced evaluations like GPQA-Diamond and AIME '26.
Technical Specifications
The model features a context window of 128,000 tokens, enabling it to process extensive datasets and long-range dependencies within a single prompt. It is designed to bridge the gap between small, efficient models and larger frontier systems by offering high intelligence at a price point typically reserved for lower-tier models. Access is primarily provided through official API endpoints and select integrated third-party model platforms.