Solar Pro 2 Preview is here: a compact 31B model that rivals 70B giants, now with hybrid reasoning mode. Still small but smarter, and free to use until July 15 on the Upstage Console.
Key Points
- 1Upstage has launched Solar Pro 2 Preview, a compact 31B parameter model designed to deliver 70B-class performance with robust reasoning capabilities and efficiency.
- 2This new model excels across English, Japanese, and Korean benchmarks, offering hybrid chat and structured reasoning modes for complex problem-solving.
- 3Key enhancements include an extended 64K token context window and improved tokenizer, with the model currently available for free preview until July 15, 2025.
Upstage has unveiled Solar Pro 2 Preview, a 31-billion-parameter (31B) language model designed to deliver high performance with reduced computational overhead, aligning with the company's philosophy of "more with less" regarding compute, GPUs, and model size.
The model is engineered to achieve performance comparable to much larger 70-billion-parameter (70B) class models, such as Llama 3.3 70B and Qwen2-72B. This efficiency is attributed to "advanced optimization and targeted pre-training," though specific technical details regarding these methodologies are not elaborated beyond this general statement. Benchmarking indicates strong multilingual capabilities, particularly excelling in English, Japanese, and Korean, where it reportedly surpasses ~70B models.
Solar Pro 2 Preview incorporates hybrid operational modes: a "chat mode" optimized for rapid responses and a "reasoning mode" for complex, multi-step problem-solving. Users can toggle between these modes using a reasoning_effort flag. The reasoning mode is designed for tasks like intricate question-answering, multi-hop queries, and agent planning. While general chat benchmarks were employed for evaluation—rather than typical reasoning benchmarks focused on mathematics, coding, or multi-hop tasks—the model's reasoning performance is stated to be on par with similarly sized models like Qwen3-32B.
Further enhancements include:
- An extended context window supporting up to 64,000 tokens ( tokens), doubling the previous limit, which facilitates the processing of significantly longer inputs and outputs.
- A tokenizer update that yields token savings ranging from 2% to 30% in specific scenarios, particularly for document-intensive or Korean-language tasks, thereby improving both processing speed and cost-efficiency.
Solar Pro 2 Preview is currently available for testing, with a full release planned that aims to introduce enhanced usability, stability for enterprise adoption, and broader deployment options through cloud marketplaces and on-premises solutions.