[K-AI Challenge] ③ Upstage Aims for 'AI Technology Leadership' with New 300B-Parameter Solar Model

Scaling Up to 300 Billion Parameters
Upstage Targets Global Frontier Performance
Consortium Unites Startups and Academia for Sustainable AI Ecosystem

Kwon Sunil, Vice President of Upstage, said, "The company's mission at the time of its founding was 'Making AI beneficial.'" He added, "Through this independent AI foundation model development project, we have the opportunity to take on challenges from a public interest perspective, beyond commercialization and productization." Photo by Park Yujin

Upstage, which was selected as one of five teams for the government's independent artificial intelligence (AI) foundation model development project, is set to develop a new AI model that will be scaled up to a final size of 300 billion parameters. This new model will be fundamentally different from Upstage's previous 'Solar' large language model (LLM), with the goal of achieving over 99% of the performance of the world's top models.

On August 13, in an interview with Asia Economy at the company's office in Yeoksam-dong, Gangnam-gu, Seoul, Upstage Vice President Kwon Sunil emphasized, "We are not simply aiming to create an AI that only works well in Korean, like a Galapagos-style isolated system. The true meaning of 'sovereign AI' is to develop a global frontier-level model and secure technological leadership." This means they intend to build a model with global competitiveness, going beyond merely improving Korean language fluency or spelling accuracy.

This philosophy is directly reflected in the composition of the consortium. Unlike other consortia led by large corporations, Upstage chose 'AI-native' startups as partners. Kwon explained, "We even considered 'work culture' in our selection process to collaborate with companies that are fully committed to AI."

Specifically, Upstage will lead the model development, Flitto will handle data collection, Nota will focus on model optimization, and Rablup will be responsible for infrastructure such as GPUs. In academia, KAIST and Sogang University will participate to support research and development (R&D). Kwon stated, "It is important to ensure the ecosystem is sustainable," adding, "During the process of forming the consortium, we proposed sharing experience and know-how through this kind of structure."

The new model to be developed by this consortium, 'Solar World Best LLM (WBL),' will take a completely different approach from the previous Solar model. Kwon explained, "Previously, there were size constraints to suit B2B (business-to-business) operations, but this time, our goal is to achieve the highest possible performance." While earlier models were limited to the range where return on investment (ROI) could be achieved for B2B business, this time they are breaking free from those constraints to pursue top-tier performance.

The new model will be developed along three main directions: achieving global frontier status in the domains of knowledge and intelligence, enhancing safety and reliability, and improving agent tool compatibility. The plan is to start with a 100 billion (100B) parameter model for the first competition at the end of the year, and then gradually expand to a final size of 300 billion (300B) parameters.

To achieve these ambitious goals, Upstage is also actively recruiting overseas talent. Among the five selected consortia, Upstage was the only one chosen for talent support, which was possible as part of this strategy. Kwon explained, "We actively sought overseas talent to secure capabilities in areas where there is little accumulated know-how in Korea or where relevant experience was gained abroad."

They plan to recruit three experts, including researchers from university labs and those with experience at big tech companies and prominent AI startups. Kwon introduced them as "well-known figures in the AI community, with extensive experience in model architecture and training." Programs are also being planned to ensure their expertise and know-how are effectively shared within the company.

Beyond model development and talent acquisition, Upstage is also emphasizing practicality. Kwon noted, "Many companies receive high benchmark scores, but feedback shows that their models do not perform as well in actual industry applications." He announced plans to create and release evaluation datasets tailored to each domain, aiming to build models that are truly usable in real-world industries, not just those with high benchmark scores.

To this end, Upstage will collaborate with consortium companies in finance, manufacturing, defense, and the public sector to jointly carry out the entire process, from collecting and validating data to evaluation, optimized for each field. Kwon explained, "Meaningful feedback comes only when models are used in industry, and this is what leads to global competitiveness."

This confidence in their strategic approach is based on the assessment that the global AI technology gap is narrowing. Regarding the government project, which aims for performance exceeding 95% of the latest models such as GPT-5, Kwon analyzed, "The algorithms and methodologies themselves have become quite standardized," and added, "Now is the time when we can catch up."

Kwon also expressed pride in Upstage's capabilities. He pointed out that about 70% of the company's 140 employees are AI R&D personnel, stating, "While our total AI workforce may be smaller than that of large corporations, if you look at the number of people solely focused on this area, it's not that small." He added, "Even before the 'ChatGPT moment'-the turning point when generative AI became mainstream and the entire AI industry entered the large language model race-we already had technology and talent in natural language processing (NLP) and natural language understanding (NLU), and we have experience in everything from fine-tuning to developing models from scratch."

With the interim evaluation scheduled for the end of the year, Kwon expressed strong confidence in being selected for the final round, saying, "Internally, we believe there is a '99% chance' of success." He added, "Since the timeline is short, our existing know-how and assets will be heavily utilized, and in that respect, we are well prepared."