Upstage Generative AI Ranks 2nd in International Competition... Demonstrating High-Level Performance

Kim Sung-hoon, CEO of Upstage [Photo by Yonhap News]

AI startup Upstage's self-developed generative AI model ranked 2nd in a global competition.

Upstage announced on the 19th that its AI model secured 2nd place on the Open LLM (Large Language Model) leaderboard operated by Hugging Face, a global AI platform.

Founded in 2020, Upstage is a domestic AI startup. It gained significant user response by introducing ‘Asukup (AskUp)’, a ‘ChatGPT with eyes’ that applies its proprietary Optical Character Recognition (OCR) technology to ChatGPT. Asukup surpassed 350,000 users within 18 days of launch and currently exceeds 1.3 million users.

Hugging Face is a global AI platform providing large language models for various languages and industries. Over 300 AI models, including mega AI models like OpenAI’s ChatGPT and Google’s Bard, compete on Hugging Face’s leaderboard. The final ranking is determined by the average evaluation scores across four indicators necessary for assessing generative AI: model inference and commonsense ability, comprehensive language understanding, and hallucination prevention.

Earlier this month, Upstage submitted its self-developed model to Hugging Face’s leaderboard for performance evaluation. As a result, Upstage’s model ranked 2nd with an average score of 64.7, following Meta’s newly released ‘LLaMA 2’. Notably, the model achieved competitive results with a size of 30B (30 billion) parameters, less than half of LLaMA 2’s 70B (70 billion) parameters.

Despite its relatively small model size, Upstage’s model scored 56.5 on the hallucination prevention metric, one of the biggest challenges in generative AI, surpassing Meta’s latest LLaMA 2 model’s score of 52.8.

Recognized for its world-class technology, Upstage’s prospects in the private AI market have also grown. Small-scale LLMs under 65B (65 billion) parameters offer significant cost savings compared to mega AI models from big tech companies, which are known to have over a trillion parameters. Additionally, they can be installed and operated on internal servers, making them highly active for private AI use by individual companies.

Private AI, which trains on data owned by individual companies to prevent internal information leaks and reduce hallucinations that generate incorrect information, is attracting attention as a new frontier in the generative AI competition. In fact, concerns about potential data leaks when entering internal company information into ChatGPT have led to increased restrictions on its use.

Upstage plans to further enhance Korean language performance by additionally training on Korean data to meet the demands of various domestic industries.

Kim Seong-hoon, CEO of Upstage, said, “We are very pleased that Upstage’s AI technology has achieved the world’s best results even in the evaluation of the most fiercely competitive global generative AI platforms. Based on the proven technology this time, we will strive to make AI conveniently accessible not only to domestic companies but to all companies worldwide.”

This content was produced with the assistance of AI translation services.