⁠Solar

Upstage AI’s Solar family of large language models (LLMs) targets enterprise use cases, with particular strength in document processing, data extraction, and multilingual understanding. These models handle vast amounts of unstructured data — they’re designed locked in business documents.

Document Parse (95%+ Accuracy) and Information Extract

Solar models deliver natural language understanding, generation, summarization, and translation. A core strength’s in its Document AI suite, which features "Document Parse" for converting diverse document formats—including PDFs, scans, and emails—into clean, machine-readable text. "Information Extract" then pulls structured key-value data from these processed documents. Beyond document intelligence, Solar’s also got conversational AI and complex reasoning capabilities.

Insurance, Healthcare, and Korean-First Enterprise

Solar’s built for enterprise use, particularly for organizations where accuracy and reliability’s critical. Specific applications include:

• Automating Workflows: Streamlining claims processing, underwriting, and policy workflows in finance and insurance.
• Clinical and Operational Analysis: Analyzing documents in healthcare for insights and efficiency.
• Customer Service: Enhancing customer interactions and providing general enterprise support.
• Custom AI Development: Building domain-specific AI applications and fine-tuning models on proprietary enterprise data.

GPT-4o-Level Reasoning with 31B Parameters and DUS Architecture

Upstage AI has engineered Solar models for high performance and efficiency:

• Reasoning Power: Solar models claim "GPT-4o Level Reasoning." Solar Pro 2 has shown comparable or superior performance to GPT-4o in benchmarks, despite its smaller size. Solar Pro 3 aims for even greater accuracy and stability in reasoning tasks.
• Context Window: Solar Pro 2 handles up to 64,000 tokens, while Solar Pro 3 expands this to 128,000 tokens, allowing for extensive document processing.
• Efficient Architecture: Techniques like Depth Up-Scaling (DUS) and Mixture-of-Experts (MoE) (for Solar Pro 3) enable high performance with relatively smaller models (e.g., Solar Pro 2 has 31 billion parameters), so you’re getting efficient deployment.
• Multilingual Prowess: The models excel in Korean, English, and Japanese, often outperforming larger models in Korean-specific benchmarks.
• Document AI Accuracy: The Document Parse tool boasts over 95% accuracy and can be up to 10 times faster than some competitors in converting documents.
• Integration Options: Solar models are accessible via API through the Upstage Console, Amazon Web Services (AWS Marketplace, SageMaker JumpStart, Bedrock), and can be deployed on-premise or in private cloud environments for stringent data privacy needs.
• Agentic Capabilities: Solar Pro 2 functions as a fully agentic LLM, capable of executing multi-step tasks by interacting with external tools like web search and presentation generation.

Korean Language Mastery and Single-GPU Deployment

Solar offers distinct advantages over more general-purpose LLMs, particularly for enterprise needs:

• Superior Document AI: It excels in processing complex, unstructured documents and accurately extracting structured data with high speed and precision.
• Korean Language Mastery: Offers superior performance and natural fluency in Korean, making it ideal for Korean-first AI services — that’s a big advantage.
• Cost-Efficient Deployment: Achieving high performance with smaller model sizes means more efficient deployment on single GPUs, reducing infrastructure costs for enterprises.
• Tailored Customization: Designed for efficient fine-tuning on private enterprise data, enabling highly specialized, domain-specific AI solutions.
• Focused Enterprise Solutions: Provides targeted, high-performance solutions for critical tasks in regulated industries, rather than broad general intelligence, and that’s the point.

Upstage Console, AWS Marketplace, and On-Premise

Solar models are available through various channels:

1. Upstage Console: Access via API through the official Upstage Console.
2. AWS Marketplace: Available on Amazon Web Services, including SageMaker JumpStart and Bedrock.
3. On-premise/Private Cloud: Deployment options for enterprises with strict data privacy requirements.

$0.15/Million Input Tokens for Solar Pro 3

Upstage AI offers a tiered pricing model for its enterprise solutions. While preview versions and educational programs may offer free access, standard enterprise use involves token-based pricing.

Upstage also provides commitment tiers and custom pricing for mission-critical workloads. Free access to Solar Pro and Document Parse is available for up to one year through AI support programs for educational and public interest organizations.

Originality Allegations, Limited Language Breadth, and Smaller Ecosystem

Several aspects of Solar warrant attention before adoption:

• Historical Allegations: In early 2026, the Solar Open 100B model faced allegations of structural similarities to a Chinese model, raising questions about its originality. Upstage publicly denied these claims.
• Japanese Language (Older Versions): An older review from March 2024 noted that Solar’s Japanese language ability was "quite low," though newer versions claim strong Japanese support. If you’ve got critical Japanese language needs, you should verify current performance.