Learning Objectives
- Understand what Solar Pro 2 is and how a 31 billion parameter model outperforms much larger competitors
- Evaluate Upstage's positioning as Korea's first potential generative AI IPO
- Compare Solar Pro 2 to frontier models on efficiency and benchmark performance
What Is Solar Pro 2?
Solar Pro 2 is a 31 billion parameter language model from South Korean AI company Upstage. Despite being significantly smaller than competitors, it scored 58 on the Artificial Analysis Intelligence Index — beating GPT-4.1 (53), DeepSeek V3 (53), and Llama 4 Maverick (51). This places Solar Pro 2 among the most efficient models in the world: achieving frontier-class performance at a fraction of the parameter count.
Released in July 2025 and updated in January 2026 with improved reasoning and instruction-following, Solar Pro 2 uses a hybrid dual-mode architecture with "Chat Mode" for fast responses and "Reasoning Mode" for multi-step logical thinking.
💡Key Concept
Small Language Model (sLLM): A category of AI models under 40 billion parameters that achieve surprisingly strong performance through architectural innovation and training efficiency — rather than raw scale. Solar Pro 2 at 31 billion parameters competes with models 3 to 10 times its size, making it ideal for cost-sensitive and edge deployments.
Key Specifications
| Spec | Detail |
|---|---|
| Parameters | 31 billion (dense) |
| Architecture | Hybrid dual-mode (Chat Mode + Reasoning Mode) |
| Languages | Strong Korean, English, and Japanese |
| License | Apache 2.0 (model weights on Hugging Face) |
| Artificial Analysis Score | 58 (beats GPT-4.1 at 53) |
| Availability | API (Upstage Console), Hugging Face, AWS Marketplace |
Benchmark Performance
| Model | Parameters | Artificial Analysis Score |
|---|---|---|
| Solar Pro 2 | 31 billion | 58 |
| GPT-4.1 | Not disclosed | 53 |
| DeepSeek V3 | 671 billion (MoE) | 53 |
| Llama 4 Maverick | 109 billion (MoE) | 51 |
| GPT-4o | Not disclosed | 41 |
Solar Pro 2 is the highest-scoring Korean-developed model on international benchmarks and is listed alongside OpenAI, Google, Meta, and Anthropic on the Artificial Analysis rankings.
📝Note
The Artificial Analysis Intelligence Index is one benchmark suite — not a universal ranking. Solar Pro 2's strong score reflects specific strengths in reasoning and instruction-following, but other benchmarks may tell a different story. Always evaluate models on your specific use case.
Company Details
| Detail | Info |
|---|---|
| Company | Upstage |
| Founded | 2020 |
| CEO | Sung Kim (former Naver Clova AI lead; ex-HKUST professor) |
| Headquarters | Yongin, South Korea (US operations in San Jose) |
| Total Funding | ~$157 million (including $72 million Series B and $45 million bridge from KDB, Amazon, AMD) |
| Valuation | ~$570 million (790 billion KRW) |
| IPO Plans | Targeting H2 2026 (Korea's first post-ChatGPT generative AI IPO); KB Securities and Mirae Asset Securities hired |
| Website | upstage.ai |
Strengths
- Exceptional efficiency — 31 billion parameters outperforming models 3 to 10 times larger on the Artificial Analysis Intelligence Index
- Dual-mode architecture — switch between fast Chat Mode and deep Reasoning Mode depending on the task
- Open source (Apache 2.0) — freely available on Hugging Face for download and self-hosting
- Strong multilingual — particularly strong in Korean, English, and Japanese
- Korean government contracts — Korea's first public-sector generative AI provider
Limitations and Considerations
- Single benchmark standout — the Artificial Analysis score is impressive but other benchmarks may show different results; evaluate on your specific tasks
- Small company — ~$157 million raised; limited capacity compared to frontier model labs
- Solar Pro 3 already announced — rapid iteration may make Solar Pro 2 superseded quickly
- IPO timeline uncertain — "as early as H2 2026" is aspirational and could slip to H1 2027
- Limited global adoption — strongest in the Korean market; less well-known internationally than Llama, Mistral, or Qwen
Key Takeaways
- Solar Pro 2 is a 31 billion parameter model that scored 58 on the Artificial Analysis Intelligence Index — beating GPT-4.1 (53) and DeepSeek V3 (53) despite being a fraction of their size
- Built by Upstage (South Korea); Apache 2.0 licensed; dual-mode architecture for fast chat and deep reasoning
- Upstage is targeting Korea's first generative AI IPO in H2 2026, with $157 million raised and Korean government contracts
- Best suited for organizations wanting frontier-class performance in a cost-efficient, self-hostable model — especially for Korean, English, and Japanese language tasks