Upstage unveiled its proprietary artificial intelligence (AI) model "Solar-Open-100B," saying it will build "the most Korean AI that understands Korean culture and even the subtle nuances of the Korean language."
Kim Sung-hoon, CEO of Upstage, said at the first briefing for the independent AI foundation model project held on the 30th at the COEX Auditorium in Samseong-dong, Seoul, "Unlike large corporations, Upstage has spent the past five years focused solely on a single goal: building AI that helps everyone."
The Solar-Open model showcased in this first release is a 100B-class (100 billion) large language model (LLM). Kim said, "This model goes beyond a simple experimental stage and has been refined to a level where it can be deployed in real services and work environments," adding, "We focused on verifying performance and efficiency at the same time."
Upstage highlighted Korean-language comprehension based on high-quality data as Solar-Open's competitive edge. Kim said, "Our goal was not merely to be proficient in Korean, but to build AI that understands context, emotion, and nuance," emphasizing distinctions between honorific and informal speech, changes in expression depending on the situation, and responses to questions that require step-by-step reasoning.
In the training process, resource efficiency and training stability were identified as core tasks. Kim said, "Because we trained with government-supported GPUs, we prioritized resource efficiency," adding, "Through automatic fault detection and failover systems and training optimization, we significantly shortened overall training time even in large-scale GPU environments."
Upstage also stressed that the model was completed through a division of roles within the consortium. The Upstage consortium is the only team among the five elite teams composed solely of startups.
Flitto provided the training and evaluation data in an end-to-end capacity. Based on its years of know-how in building and operating multilingual data, Flitto designed high-difficulty training data, real-use evaluation data, and scenario-based quality verification data, and supported model training with data that reflects the cultural background and context of the Korean language.
Meanwhile, Lablup handled operation of the large-scale training infrastructure. Kim said, "In environments where thousands of GPUs are run simultaneously, even a minor fault can halt the entire training," adding, "In collaboration with Lablup, we built a system that automatically detects failures and provides immediate failover."
Real-world validation of the model involved organizations across various domains, including Law&Company, MakinaRocks, and VUNO. Kim said, "Each corporation tested Solar-Open-100B in actual work environments to verify its usability, and this feedback played an important role in improving the model's completeness."
Upstage aims not to keep Solar-Open-100B as a closed model for use by specific corporations only, but to spread it through diverse use cases. Kim said the consortium is also preparing applied services using Solar-Open, introducing Day 1 Company's planned "nationwide hackathon" as a representative example. Kim added that the technical know-how accumulated during model development is being shared through 10 lectures, with six completed so far.
Going forward, the company plans to broaden applications to agent-style use that executes users' goals in search, fact-checking, research, and document drafting.
Kim emphasized, "The AI we set out to build is not just a model that speaks well, but an AI that actually works well," adding, "Rather than chasing specific global corporations, we will build AI that can compete in the global market with independent technology."
Upstage also unveiled plans for the next model. Next year, it will scale up to a 200B-class LLM, train on 15 trillion tokens, support a 256K context, and add three languages: Korean, English, and Japanese. Ultimately, with a focus on generalization and popularization, the company plans to expand the model to the 300B class while broadening the scope of training data and context.
Kim added, "Our goal is one," saying, "Together with many consortia, we will help build a global AI Big Three that can compete with Google and OpenAI using the Solar LLM."
In January next year, the government will conduct a first-stage assessment to comprehensively review the performance and future plans of the teams participating in the independent AI foundation model project, and will narrow the elite teams down to four based on the results. After that, reviews will be held every six months to reduce the number of elite teams by one each time, with two teams to be finally selected in 2027.