Vibeleaderboard
Upstage

ai lab

Upstage

Upstage matters because it tests whether a regionally grounded lab can combine competitive local-language models with a specific enterprise workflow instead of copying the broad assistant strategy of larger labs. SOLAR also supplied open researchers with a simple depth-scaling method, while Solar Pro 4 now gives the company independently measured evidence that its proprietary line is improving.3,6,14,5

Official site ↗Models in the Atlas →

Profile

Overview

A team formed from Naver Clova

Upstage is a South Korean AI company founded in 2020 by Sunghun Kim, Hwalsuk Lee, and Lucy Park. All three came from senior AI roles at Naver: Kim led Clova AI, Lee led Clova Visual AI and optical character recognition work, and Park led modeling for the Papago translation service. The company therefore began with two connected specialties, language models and systems that turn business documents into usable data.1,2

SOLAR and depth up-scaling

Its first widely examined model was SOLAR 10.7B. The associated paper introduced depth up-scaling, a method that expands a smaller pretrained network by duplicating and rearranging layers before continued pretraining. Upstage released the weights under Apache 2.0 and published the work at the NAACL industry track. The contribution was an efficient scaling recipe and a reproducible model artifact, not the invention of compact language models.3,4

From a compact open model to Solar Pro

Upstage developed Solar into a commercial family aimed at Korean, English, and enterprise workloads. Upstage entered South Korea's sovereign-model program with Solar Pro 2 as its public model reference, Solar Pro 3 moved to a sparse mixture-of-experts design and added stronger reasoning and tool-use training, and Solar Pro 4 focused on long-context and multi-step office work. Artificial Analysis measured a large gain from Solar Pro 3 to Solar Pro 4, while still placing the newer model below the top frontier systems on its overall index.6,10,11,14

Documents remain the commercial center

The other half of the company is Document AI. Document Parse converts complex pages into structured text for retrieval and language-model workflows, while Upstage Studio packages parsing, classification, extraction, and agentic document processes for enterprises. Funding from Korean institutions, Amazon, and AMD has helped Upstage expand this combined model-and-document strategy. The result is a focused enterprise lab, not a general consumer assistant company.5,7,8

Notable contributions

  1. 01Depth up-scaling for SOLARUpstage introduced and documented a practical method for expanding a pretrained transformer by changing its depth and continuing training. The claim is limited to this published recipe, not model scaling in general.3
  2. 02A Korean-focused compact model programThe Solar family made Korean-language quality, serving efficiency, and enterprise deployment first-class design constraints across open and proprietary releases.6,10,14
  3. 03Document parsing as model infrastructureUpstage treats page reconstruction, tables, layout, classification, and extraction as infrastructure that precedes language-model reasoning, then packages those capabilities with its models in one enterprise stack.5,2