Upstage is a South Korean artificial intelligence company that develops large language models and document processing technologies. Founded in 2020, the company focuses on building enterprise-grade Generative AI solutions, including its Solar series of large language models and document AI tools. Upstage gained international recognition when its models appeared on the LMArena leaderboard, a public platform where users compare and rank AI systems through blind testing.
The company operates from its headquarters in Pangyo, South Korea, a major technology hub south of Seoul. Upstage has positioned itself as a bridge between advanced Artificial intelligence research and practical business applications, particularly for organizations in Asia and global markets. Its work spans Machine learning research, Deep learning architectures, and applied AI services.
Founding and Early Development
Upstage was established in October 2020 by a team of engineers and researchers with backgrounds in major technology firms and academic institutions. The founding team included individuals who had previously worked at companies such as Samsung Electronics and Google DeepMind, bringing experience in both hardware-adjacent software and frontier AI research. The company initially focused on document AI, developing systems that could extract and structure information from unstructured documents, a common challenge in industries like finance, legal, and healthcare.
In its first two years, Upstage built proprietary Transformer (architecture)-based models for optical character recognition and document understanding. These early products attracted enterprise clients in South Korea, including banks and government agencies, establishing a revenue base that funded subsequent research into larger generative models.
Solar Language Models
In 2023, Upstage released Solar, its first family of large language models. The initial Solar model, Solar 10.7B, was notable for its parameter efficiency - the 10.7 billion parameter model was designed to compete with larger systems while requiring less computational resources. This efficiency was achieved through architectural innovations in the Neural network design, including a method the company called depth-up scaling, which involved stacking layers from smaller pretrained models.
Solar models were trained on multilingual datasets with a strong emphasis on Korean and English, reflecting Upstage's dual focus on domestic and international markets. The models were made available through open-source licenses on platforms like Hugging Face, allowing developers to fine-tune them for specific tasks. By early 2024, Upstage had released subsequent versions, including Solar Pro, a larger model aimed at enterprise deployment.
The company's participation in LMArena, a crowdsourced benchmarking platform, helped raise its global profile. Solar models consistently ranked among the top performers in their size class, competing with systems from larger organizations like OpenAI and Anthropic. This visibility led to partnerships with cloud providers and AI infrastructure companies.
Document AI and Enterprise Services
Beyond language models, Upstage maintains a significant document AI business. Its platform offers services such as document parsing, table extraction, and question-answering over proprietary documents. These tools are designed to integrate with existing enterprise workflows, often deployed through Amazon Web Services or Microsoft Azure cloud environments.
The company targets industries with heavy document processing needs, including insurance, banking, and public administration. Upstage reports that its document AI systems can process millions of pages per month for clients, with accuracy rates that improve through continuous fine-tuning on customer-specific data. This dual focus on generative models and structured document processing distinguishes Upstage from many competitors that concentrate solely on chat-based AI.
Research and Open Source Contributions
Upstage maintains an active research division that publishes papers on topics such as model efficiency, multilingual training, and alignment techniques. The company has contributed to open-source projects, including fine-tuning frameworks and evaluation tools used by the broader Machine learning community. Its researchers have collaborated with academic institutions, including University of Toronto and Carnegie Mellon University, on joint studies.
The company also participates in industry consortia and standards efforts related to AI safety and evaluation. Upstage has advocated for transparent benchmarking, supporting initiatives that allow independent verification of model capabilities. Its presence on LMArena is part of this philosophy, as the platform provides real-time, user-driven assessments rather than static benchmarks.
Market Position and Future Direction
Upstage operates in a competitive landscape that includes both global giants and regional startups. While it lacks the scale of OpenAI or Google DeepMind, its focus on efficiency and enterprise integration has carved a niche. The company has raised significant venture funding, with investors including South Korean conglomerates and international firms, valuing it in the hundreds of millions of dollars as of 2024.
Looking forward, Upstage plans to expand its multilingual capabilities beyond Korean and English, targeting Japanese and Southeast Asian markets. It is also exploring on-device AI applications, potentially partnering with hardware makers like Qualcomm or Arm Holdings to deploy smaller models on edge devices. The company continues to refine its Solar architecture, aiming to achieve frontier-level performance with fewer parameters, a goal that aligns with broader industry trends toward efficiency in Deep learning.