Solar Pro 2

Upstage has officially launched its next-generation LLM, Solar Pro 2. It has grown from 22B parameters in the previous generation to 31B, but still falls within the sLLM category.

The model offers two user-selectable modes.

• Chat Mode: focused on fast responses
• Reasoning Mode: focused on structured, multi-step logical reasoning

In particular, through Chain-of-Thought (CoT) reasoning, it has improved performance on math, coding, and logic-centered tasks.

In benchmarks, the company said it achieved results comparable to larger models such as GPT-4o, DeepSeek R1, Mistral Small 3.2, Qwen 3 on MMLU-Pro, Math500, AIME, SWE-Bench, and others. Based on the fact that this performance was shown at less than half the size, Upstage emphasizes that it has proven global competitiveness entirely with its own technology.

Korean language ability was also highlighted as a strength. On Arena-Hard-Auto, Hae-Rae, Ko-MMLU, it showed results comparable to or better than competing models, and it explained that it is strong in vocabulary, context understanding, and cultural nuance.

Also, Solar Pro 2 has evolved beyond a simple conversational LLM into an agentic LLM that performs tasks by linking with external tools. For example, it can automatically research competitors' recent product launches and marketing activities and even produce a presentation with attached links, handling multi-step work from planning through to producing the final output.

It is currently available via the API on the Upstage Console.