Solar Pro 4 - LLM Specialized for Agentic Work

Upstage has released Solar Pro 4, an agent-specialized LLM that reads documents, executes tools, and completes final artifacts. Designed to go beyond answering questions, it performs multi-step tasks such as contract review, cross-file numerical comparison, and result verification in sequence.

Solar Pro 4 supports a 512K context and up to 128K output, with input and output capabilities in English, Korean, and Japanese. The reasoning effort setting allows users to adjust reasoning depth and response speed, expanding the scope of use from complex analysis to real-time interaction.

Agent evaluation performance has significantly improved compared to Solar Pro 3.

• Terminal-Bench v2.1: 12 points → 57 points
• τ³-Banking: 9 points → 23 points
• AA-LCR: 31 points → 71 points

The improvements leveraged OfficeVerse, a proprietary pipeline that synthesizes and evaluates 11 industry domains and 12 office work types based on actual public data. The model was trained and validated using workflows close to real-world tasks, such as Excel analysis and PowerPoint presentations, and included Ko-GDPval, a Korean office work evaluation.

While Solar Open 2 is a general-purpose open-weight model suitable for on-premises operations, Solar Pro 4 is a model that handles commercial agent tasks requiring multiple documents, terminal operations, and multi-stage tool calls via API. The difference between the two models was particularly pronounced in the agent execution domain, with Solar Pro 4 recording 57.0 points on Terminal-Bench, 49.2 points on BrowseComp, and 38.8 points on GDPval-AA v2.

In a real-world case, the model evaluated 10 candidate locations based on one regulatory document and six commercial district data files, creating an Excel workbook, a review report, and a PowerPoint presentation with just three prompts. The calculation results in the workbook were linked to the report and presentation to verify numerical consistency across artifacts.

Additionally, for questions without basis in the documents, the model applies a method of answering that it cannot verify rather than guessing the content. Upstage also released the Solar Pro 4 Agent Cookbook, which includes agent implementation examples based on Excel, Word, and PowerPoint, along with system prompts, output results, and passing criteria.