Solar Pro 2 Preview: Small. Powerful. Now with reasoning.

Enterprises are always looking for better performance at lower cost. At Upstage, we’ve consistently focused on delivering more with less—less compute, fewer GPUs, and smaller models that outperform their size. That philosophy powered Solar Mini and Solar Pro.

Today, we’re excited to unveil Solar Pro 2 Preview—our most capable model yet. It’s compact and efficient as always, now with robust reasoning capabilities. Here’s what sets it apart:

• Small yet powerful: At 31B parameters, Solar Pro 2 Preview matches the performance of models up to 70B in size.
• Multilingual strength: It excels in English, Japanese, and especially Korean—outperforming ~70B models in multiple benchmarks.
• Hybrid modes: Users can choose between chat mode for fast responses and reasoning mode for structured, multi-step problem solving.

31B model with 72B-class performance

Solar Pro 2 Preview rivals models twice its size, like Llama 3.3 70B and Qwen2-72B. Through advanced optimization and targeted pre-training, it delivers top-tier performance while remaining deployable at a fraction of the cost.

The table below compares Solar Pro 2 Preview with leading open-source models across English, Korean, and Japanese benchmarks. It includes general and reasoning tasks, highlighting our model’s competitive edge despite its smaller size.

Reasoning mode for complex tasks

Reasoning mode enables structured, multi-step thinking for tasks like complex QA, multi-hop queries, and agent planning. With the reasoning_effort flag, users can seamlessly switch between deep reasoning and casual conversation. As can be seen in the comparison table above, Solar Pro 2 Preview delivers reasoning performance on par with similarly sized models like Qwen3-32B.

Note that while reasoning is typically evaluated using benchmarks focused on math, coding, and multi-hop tasks, this comparison uses general chat benchmarks instead, due to the broader task coverage and alignment with enterprise use cases.

Additional enhancements that matter

Alongside major upgrades, Solar Pro 2 Preview includes key improvements that boost usability and performance:

• Extended context window: Now supports up to 64K tokens, doubling the previous 32K limit. This enables handling much longer inputs and outputs in a single pass.
• Tokenizer update: Improved tokenization leads to token savings of 2–30% in certain scenarios—especially in document-heavy or Korean-language tasks—enhancing both speed and cost-efficiency.

What’s Next?

Solar Pro 2 Preview is available now and free to use until July 15th, 2025, on the Upstage Console. We’re already preparing for the full release of Solar Pro 2, which will bring further enhancements, such as:

• Improved usability and stability for enterprise adoption
• Enterprise-grade deployment through cloud marketplaces and on-premises

Start testing with a smaller, smarter, and more reasoning-capable model—while you wait for what’s next.

‍