Upstage's proprietary agentic flagship, built for multi-step work across long documents, terminal tasks, and tool calls. It supports a 512K-token context window, up to 128K output tokens, adjustable reasoning effort, and English, Korean, and Japanese. Upstage says it trained and validated the model with OfficeVerse, a pipeline that synthesizes and grades complete office-work deliverables across 11 industry domains and 12 task types.

Compared with Solar Pro 3, Upstage reports the largest gains on work-like agent evaluations, including +13.8 points on terminal tasks, +11.9 on BrowseComp web research, and +7.4 on GDPval. The model is API-only; its architecture, parameter count, training scale, and base checkpoint have not been disclosed.

Model Details

Context window 524,288
AA Intelligence 27 on v4.3.2 was 28 on v4.3.2
License Proprietary
enterpriseagenticreasoningmultilingual

Related